|
mcap_anonymizer package from mcap_anonymizer repomcap_anonymizer |
ROS Distro
|
Package Summary
| Version | 0.1.0 |
| License | BSD-3-Clause |
| Build type | AMENT_PYTHON |
| Use | RECOMMENDED |
Repository Summary
| Checkout URI | https://github.com/MOLAorg/mcap_anonymizer.git |
| VCS Type | git |
| VCS Version | main |
| Last Updated | 2026-11-18 |
| Dev Status | DEVELOPED |
| Released | UNRELEASED |
| Contributing |
Help Wanted (-)
Good First Issues (-) Pull Requests to Review (-) |
Package Description
Additional Links
Maintainers
- Jose Luis Blanco-Claraco
Authors
- Jose Luis Blanco-Claraco
mcap_anonymizer
| Distro | Build dev | Release |
|---|---|---|
| ROS 2 Humble (u22.04) | ||
| ROS 2 Jazzy (u24.04) | ||
| ROS 2 Kilted (u24.04) | ||
| ROS 2 Lyrical (u26.04) | ||
| ROS 2 Rolling (u26.04) |
Blurs the faces in the camera images of MCAP files (ROS 2 bags), so datasets recorded among people can be shared without privacy issues, while keeping everything else exactly as it was recorded.
-
Only the faces change. In raw images (
sensor_msgs/Image), every pixel outside the blurred faces stays bit-exact, and so do the encoding,step, header and message size. Each encoding is processed in its own format: a Bayer image is still a valid mosaic, a YUV 4:2:2 image keeps its chroma subsampling, 16-bit images keep their depth and byte order. -
Everything else is untouched. All other messages are copied byte by byte, with their log
and publish times and sequence numbers, in the same order. Schemas, channels (also those
without messages, with their QoS metadata), metadata and attachment records, and the chunk
compression are kept. For a rosbag2 directory, its
metadata.yamlis copied. -
Compressed images (
sensor_msgs/CompressedImage): PNG is re-encoded losslessly (same channels and bit depth); JPEG is re-encoded with its own quantization tables and chroma subsampling, so the extra loss in the rest of the image is minimal. - Privacy first. A low detection threshold, boxes enlarged by a margin and propagated to the neighboring frames (so a face missed by the detector in a few frames is still covered), a review video to check the result, and hand-added boxes for anything missed. Topics that may contain images but cannot be processed (e.g. compressed video) stop the tool, instead of being silently copied.
- Large and unfinished bags. Files are streamed (memory use is about one MCAP chunk, whatever the bag size), and recordings that were never closed (no summary section, truncated last chunk) are processed up to their last complete record. The output is always a complete, indexed MCAP file.
Faces are detected with CenterFace at a low threshold (high recall), and doubtful detections are kept only if a person detector (YOLOX-s) finds a person around them, which removes most false positives (foliage, wheels, machinery…). Both are ONNX models that run with OpenCV (no extra dependencies), or with onnxruntime on the CPU or an NVIDIA GPU. All the bundled code and models have permissive licenses (BSD, MIT, Apache-2.0).
Install
ROS 2 (Humble, Jazzy, Kilted, Lyrical, Rolling)
From source, in a colcon workspace:
cd ~/ros2_ws/src
git clone https://github.com/MOLAorg/mcap_anonymizer.git
cd ~/ros2_ws
rosdep install --from-paths src --ignore-src -y
colcon build --packages-select mcap_anonymizer
source install/setup.bash
ros2 run mcap_anonymizer mcap_anonymize --help
Its only dependencies are python3-numpy, python3-opencv and python3-pil; the MCAP reader
and writer (the mcap Python library, MIT) is bundled.
pip (no ROS needed)
pip install 'mcap-anonymizer[compression] @ git+https://github.com/MOLAorg/mcap_anonymizer.git'
mcap_anonymize --help
Optional extras
-
zstd or lz4 compressed MCAP chunks need the Python packages
zstandardorlz4(Ubuntu:sudo apt install python3-zstandard python3-lz4; pip: the[compression]extra). Bags recorded with rosbag2’s default MCAP settings are not compressed. -
Faster detection with onnxruntime:
pip install onnxruntime. It is used automatically when installed (--backend auto). -
NVIDIA GPU:
pip install 'onnxruntime-gpu[cuda,cudnn]'(and notonnxruntime), then--device cuda(orauto). Older GPUs (Pascal: GTX 10xx, Quadro Pxxx) are not supported by the CUDA 13 libraries of the latest versions; on them, usepip install 'onnxruntime-gpu[cuda,cudnn]==1.22.0' 'nvidia-cudnn-cu12==9.5.1.17'.
Quick start
mcap_anonymize run input.mcap output.mcap --detections detections.json
mcap_anonymize run my_rosbag2_dir/ my_rosbag2_dir_anonymized/ # all the .mcap files of a bag
With ROS 2, use ros2 run mcap_anonymizer mcap_anonymize instead of mcap_anonymize.
All the image topics are processed. The tool prints how many images of each topic were
anonymized, and detections.json keeps the detected faces.
Recommended workflow to publish a dataset
```bash
1. Detect the faces (the slow part; see “Performance”)
mcap_anonymize detect input.mcap detections.json
File truncated at 100 lines see the full file
Package Dependencies
System Dependencies
Dependant Packages
Launch files
Messages
Services
Plugins
Recent questions tagged mcap_anonymizer at Robotics Stack Exchange
|
mcap_anonymizer package from mcap_anonymizer repomcap_anonymizer |
ROS Distro
|
Package Summary
| Version | 0.1.0 |
| License | BSD-3-Clause |
| Build type | AMENT_PYTHON |
| Use | RECOMMENDED |
Repository Summary
| Checkout URI | https://github.com/MOLAorg/mcap_anonymizer.git |
| VCS Type | git |
| VCS Version | main |
| Last Updated | 2026-11-18 |
| Dev Status | DEVELOPED |
| Released | UNRELEASED |
| Contributing |
Help Wanted (-)
Good First Issues (-) Pull Requests to Review (-) |
Package Description
Additional Links
Maintainers
- Jose Luis Blanco-Claraco
Authors
- Jose Luis Blanco-Claraco
mcap_anonymizer
| Distro | Build dev | Release |
|---|---|---|
| ROS 2 Humble (u22.04) | ||
| ROS 2 Jazzy (u24.04) | ||
| ROS 2 Kilted (u24.04) | ||
| ROS 2 Lyrical (u26.04) | ||
| ROS 2 Rolling (u26.04) |
Blurs the faces in the camera images of MCAP files (ROS 2 bags), so datasets recorded among people can be shared without privacy issues, while keeping everything else exactly as it was recorded.
-
Only the faces change. In raw images (
sensor_msgs/Image), every pixel outside the blurred faces stays bit-exact, and so do the encoding,step, header and message size. Each encoding is processed in its own format: a Bayer image is still a valid mosaic, a YUV 4:2:2 image keeps its chroma subsampling, 16-bit images keep their depth and byte order. -
Everything else is untouched. All other messages are copied byte by byte, with their log
and publish times and sequence numbers, in the same order. Schemas, channels (also those
without messages, with their QoS metadata), metadata and attachment records, and the chunk
compression are kept. For a rosbag2 directory, its
metadata.yamlis copied. -
Compressed images (
sensor_msgs/CompressedImage): PNG is re-encoded losslessly (same channels and bit depth); JPEG is re-encoded with its own quantization tables and chroma subsampling, so the extra loss in the rest of the image is minimal. - Privacy first. A low detection threshold, boxes enlarged by a margin and propagated to the neighboring frames (so a face missed by the detector in a few frames is still covered), a review video to check the result, and hand-added boxes for anything missed. Topics that may contain images but cannot be processed (e.g. compressed video) stop the tool, instead of being silently copied.
- Large and unfinished bags. Files are streamed (memory use is about one MCAP chunk, whatever the bag size), and recordings that were never closed (no summary section, truncated last chunk) are processed up to their last complete record. The output is always a complete, indexed MCAP file.
Faces are detected with CenterFace at a low threshold (high recall), and doubtful detections are kept only if a person detector (YOLOX-s) finds a person around them, which removes most false positives (foliage, wheels, machinery…). Both are ONNX models that run with OpenCV (no extra dependencies), or with onnxruntime on the CPU or an NVIDIA GPU. All the bundled code and models have permissive licenses (BSD, MIT, Apache-2.0).
Install
ROS 2 (Humble, Jazzy, Kilted, Lyrical, Rolling)
From source, in a colcon workspace:
cd ~/ros2_ws/src
git clone https://github.com/MOLAorg/mcap_anonymizer.git
cd ~/ros2_ws
rosdep install --from-paths src --ignore-src -y
colcon build --packages-select mcap_anonymizer
source install/setup.bash
ros2 run mcap_anonymizer mcap_anonymize --help
Its only dependencies are python3-numpy, python3-opencv and python3-pil; the MCAP reader
and writer (the mcap Python library, MIT) is bundled.
pip (no ROS needed)
pip install 'mcap-anonymizer[compression] @ git+https://github.com/MOLAorg/mcap_anonymizer.git'
mcap_anonymize --help
Optional extras
-
zstd or lz4 compressed MCAP chunks need the Python packages
zstandardorlz4(Ubuntu:sudo apt install python3-zstandard python3-lz4; pip: the[compression]extra). Bags recorded with rosbag2’s default MCAP settings are not compressed. -
Faster detection with onnxruntime:
pip install onnxruntime. It is used automatically when installed (--backend auto). -
NVIDIA GPU:
pip install 'onnxruntime-gpu[cuda,cudnn]'(and notonnxruntime), then--device cuda(orauto). Older GPUs (Pascal: GTX 10xx, Quadro Pxxx) are not supported by the CUDA 13 libraries of the latest versions; on them, usepip install 'onnxruntime-gpu[cuda,cudnn]==1.22.0' 'nvidia-cudnn-cu12==9.5.1.17'.
Quick start
mcap_anonymize run input.mcap output.mcap --detections detections.json
mcap_anonymize run my_rosbag2_dir/ my_rosbag2_dir_anonymized/ # all the .mcap files of a bag
With ROS 2, use ros2 run mcap_anonymizer mcap_anonymize instead of mcap_anonymize.
All the image topics are processed. The tool prints how many images of each topic were
anonymized, and detections.json keeps the detected faces.
Recommended workflow to publish a dataset
```bash
1. Detect the faces (the slow part; see “Performance”)
mcap_anonymize detect input.mcap detections.json
File truncated at 100 lines see the full file
Package Dependencies
System Dependencies
Dependant Packages
Launch files
Messages
Services
Plugins
Recent questions tagged mcap_anonymizer at Robotics Stack Exchange
|
mcap_anonymizer package from mcap_anonymizer repomcap_anonymizer |
ROS Distro
|
Package Summary
| Version | 0.1.0 |
| License | BSD-3-Clause |
| Build type | AMENT_PYTHON |
| Use | RECOMMENDED |
Repository Summary
| Checkout URI | https://github.com/MOLAorg/mcap_anonymizer.git |
| VCS Type | git |
| VCS Version | main |
| Last Updated | 2026-11-18 |
| Dev Status | DEVELOPED |
| Released | UNRELEASED |
| Contributing |
Help Wanted (-)
Good First Issues (-) Pull Requests to Review (-) |
Package Description
Additional Links
Maintainers
- Jose Luis Blanco-Claraco
Authors
- Jose Luis Blanco-Claraco
mcap_anonymizer
| Distro | Build dev | Release |
|---|---|---|
| ROS 2 Humble (u22.04) | ||
| ROS 2 Jazzy (u24.04) | ||
| ROS 2 Kilted (u24.04) | ||
| ROS 2 Lyrical (u26.04) | ||
| ROS 2 Rolling (u26.04) |
Blurs the faces in the camera images of MCAP files (ROS 2 bags), so datasets recorded among people can be shared without privacy issues, while keeping everything else exactly as it was recorded.
-
Only the faces change. In raw images (
sensor_msgs/Image), every pixel outside the blurred faces stays bit-exact, and so do the encoding,step, header and message size. Each encoding is processed in its own format: a Bayer image is still a valid mosaic, a YUV 4:2:2 image keeps its chroma subsampling, 16-bit images keep their depth and byte order. -
Everything else is untouched. All other messages are copied byte by byte, with their log
and publish times and sequence numbers, in the same order. Schemas, channels (also those
without messages, with their QoS metadata), metadata and attachment records, and the chunk
compression are kept. For a rosbag2 directory, its
metadata.yamlis copied. -
Compressed images (
sensor_msgs/CompressedImage): PNG is re-encoded losslessly (same channels and bit depth); JPEG is re-encoded with its own quantization tables and chroma subsampling, so the extra loss in the rest of the image is minimal. - Privacy first. A low detection threshold, boxes enlarged by a margin and propagated to the neighboring frames (so a face missed by the detector in a few frames is still covered), a review video to check the result, and hand-added boxes for anything missed. Topics that may contain images but cannot be processed (e.g. compressed video) stop the tool, instead of being silently copied.
- Large and unfinished bags. Files are streamed (memory use is about one MCAP chunk, whatever the bag size), and recordings that were never closed (no summary section, truncated last chunk) are processed up to their last complete record. The output is always a complete, indexed MCAP file.
Faces are detected with CenterFace at a low threshold (high recall), and doubtful detections are kept only if a person detector (YOLOX-s) finds a person around them, which removes most false positives (foliage, wheels, machinery…). Both are ONNX models that run with OpenCV (no extra dependencies), or with onnxruntime on the CPU or an NVIDIA GPU. All the bundled code and models have permissive licenses (BSD, MIT, Apache-2.0).
Install
ROS 2 (Humble, Jazzy, Kilted, Lyrical, Rolling)
From source, in a colcon workspace:
cd ~/ros2_ws/src
git clone https://github.com/MOLAorg/mcap_anonymizer.git
cd ~/ros2_ws
rosdep install --from-paths src --ignore-src -y
colcon build --packages-select mcap_anonymizer
source install/setup.bash
ros2 run mcap_anonymizer mcap_anonymize --help
Its only dependencies are python3-numpy, python3-opencv and python3-pil; the MCAP reader
and writer (the mcap Python library, MIT) is bundled.
pip (no ROS needed)
pip install 'mcap-anonymizer[compression] @ git+https://github.com/MOLAorg/mcap_anonymizer.git'
mcap_anonymize --help
Optional extras
-
zstd or lz4 compressed MCAP chunks need the Python packages
zstandardorlz4(Ubuntu:sudo apt install python3-zstandard python3-lz4; pip: the[compression]extra). Bags recorded with rosbag2’s default MCAP settings are not compressed. -
Faster detection with onnxruntime:
pip install onnxruntime. It is used automatically when installed (--backend auto). -
NVIDIA GPU:
pip install 'onnxruntime-gpu[cuda,cudnn]'(and notonnxruntime), then--device cuda(orauto). Older GPUs (Pascal: GTX 10xx, Quadro Pxxx) are not supported by the CUDA 13 libraries of the latest versions; on them, usepip install 'onnxruntime-gpu[cuda,cudnn]==1.22.0' 'nvidia-cudnn-cu12==9.5.1.17'.
Quick start
mcap_anonymize run input.mcap output.mcap --detections detections.json
mcap_anonymize run my_rosbag2_dir/ my_rosbag2_dir_anonymized/ # all the .mcap files of a bag
With ROS 2, use ros2 run mcap_anonymizer mcap_anonymize instead of mcap_anonymize.
All the image topics are processed. The tool prints how many images of each topic were
anonymized, and detections.json keeps the detected faces.
Recommended workflow to publish a dataset
```bash
1. Detect the faces (the slow part; see “Performance”)
mcap_anonymize detect input.mcap detections.json
File truncated at 100 lines see the full file
Package Dependencies
System Dependencies
Dependant Packages
Launch files
Messages
Services
Plugins
Recent questions tagged mcap_anonymizer at Robotics Stack Exchange
|
mcap_anonymizer package from mcap_anonymizer repomcap_anonymizer |
ROS Distro
|
Package Summary
| Version | 0.1.0 |
| License | BSD-3-Clause |
| Build type | AMENT_PYTHON |
| Use | RECOMMENDED |
Repository Summary
| Checkout URI | https://github.com/MOLAorg/mcap_anonymizer.git |
| VCS Type | git |
| VCS Version | main |
| Last Updated | 2026-11-18 |
| Dev Status | DEVELOPED |
| Released | UNRELEASED |
| Contributing |
Help Wanted (-)
Good First Issues (-) Pull Requests to Review (-) |
Package Description
Additional Links
Maintainers
- Jose Luis Blanco-Claraco
Authors
- Jose Luis Blanco-Claraco
mcap_anonymizer
| Distro | Build dev | Release |
|---|---|---|
| ROS 2 Humble (u22.04) | ||
| ROS 2 Jazzy (u24.04) | ||
| ROS 2 Kilted (u24.04) | ||
| ROS 2 Lyrical (u26.04) | ||
| ROS 2 Rolling (u26.04) |
Blurs the faces in the camera images of MCAP files (ROS 2 bags), so datasets recorded among people can be shared without privacy issues, while keeping everything else exactly as it was recorded.
-
Only the faces change. In raw images (
sensor_msgs/Image), every pixel outside the blurred faces stays bit-exact, and so do the encoding,step, header and message size. Each encoding is processed in its own format: a Bayer image is still a valid mosaic, a YUV 4:2:2 image keeps its chroma subsampling, 16-bit images keep their depth and byte order. -
Everything else is untouched. All other messages are copied byte by byte, with their log
and publish times and sequence numbers, in the same order. Schemas, channels (also those
without messages, with their QoS metadata), metadata and attachment records, and the chunk
compression are kept. For a rosbag2 directory, its
metadata.yamlis copied. -
Compressed images (
sensor_msgs/CompressedImage): PNG is re-encoded losslessly (same channels and bit depth); JPEG is re-encoded with its own quantization tables and chroma subsampling, so the extra loss in the rest of the image is minimal. - Privacy first. A low detection threshold, boxes enlarged by a margin and propagated to the neighboring frames (so a face missed by the detector in a few frames is still covered), a review video to check the result, and hand-added boxes for anything missed. Topics that may contain images but cannot be processed (e.g. compressed video) stop the tool, instead of being silently copied.
- Large and unfinished bags. Files are streamed (memory use is about one MCAP chunk, whatever the bag size), and recordings that were never closed (no summary section, truncated last chunk) are processed up to their last complete record. The output is always a complete, indexed MCAP file.
Faces are detected with CenterFace at a low threshold (high recall), and doubtful detections are kept only if a person detector (YOLOX-s) finds a person around them, which removes most false positives (foliage, wheels, machinery…). Both are ONNX models that run with OpenCV (no extra dependencies), or with onnxruntime on the CPU or an NVIDIA GPU. All the bundled code and models have permissive licenses (BSD, MIT, Apache-2.0).
Install
ROS 2 (Humble, Jazzy, Kilted, Lyrical, Rolling)
From source, in a colcon workspace:
cd ~/ros2_ws/src
git clone https://github.com/MOLAorg/mcap_anonymizer.git
cd ~/ros2_ws
rosdep install --from-paths src --ignore-src -y
colcon build --packages-select mcap_anonymizer
source install/setup.bash
ros2 run mcap_anonymizer mcap_anonymize --help
Its only dependencies are python3-numpy, python3-opencv and python3-pil; the MCAP reader
and writer (the mcap Python library, MIT) is bundled.
pip (no ROS needed)
pip install 'mcap-anonymizer[compression] @ git+https://github.com/MOLAorg/mcap_anonymizer.git'
mcap_anonymize --help
Optional extras
-
zstd or lz4 compressed MCAP chunks need the Python packages
zstandardorlz4(Ubuntu:sudo apt install python3-zstandard python3-lz4; pip: the[compression]extra). Bags recorded with rosbag2’s default MCAP settings are not compressed. -
Faster detection with onnxruntime:
pip install onnxruntime. It is used automatically when installed (--backend auto). -
NVIDIA GPU:
pip install 'onnxruntime-gpu[cuda,cudnn]'(and notonnxruntime), then--device cuda(orauto). Older GPUs (Pascal: GTX 10xx, Quadro Pxxx) are not supported by the CUDA 13 libraries of the latest versions; on them, usepip install 'onnxruntime-gpu[cuda,cudnn]==1.22.0' 'nvidia-cudnn-cu12==9.5.1.17'.
Quick start
mcap_anonymize run input.mcap output.mcap --detections detections.json
mcap_anonymize run my_rosbag2_dir/ my_rosbag2_dir_anonymized/ # all the .mcap files of a bag
With ROS 2, use ros2 run mcap_anonymizer mcap_anonymize instead of mcap_anonymize.
All the image topics are processed. The tool prints how many images of each topic were
anonymized, and detections.json keeps the detected faces.
Recommended workflow to publish a dataset
```bash
1. Detect the faces (the slow part; see “Performance”)
mcap_anonymize detect input.mcap detections.json
File truncated at 100 lines see the full file
Package Dependencies
System Dependencies
Dependant Packages
Launch files
Messages
Services
Plugins
Recent questions tagged mcap_anonymizer at Robotics Stack Exchange
|
mcap_anonymizer package from mcap_anonymizer repomcap_anonymizer |
ROS Distro
|
Package Summary
| Version | 0.1.0 |
| License | BSD-3-Clause |
| Build type | AMENT_PYTHON |
| Use | RECOMMENDED |
Repository Summary
| Checkout URI | https://github.com/MOLAorg/mcap_anonymizer.git |
| VCS Type | git |
| VCS Version | main |
| Last Updated | 2026-11-18 |
| Dev Status | DEVELOPED |
| Released | UNRELEASED |
| Contributing |
Help Wanted (-)
Good First Issues (-) Pull Requests to Review (-) |
Package Description
Additional Links
Maintainers
- Jose Luis Blanco-Claraco
Authors
- Jose Luis Blanco-Claraco
mcap_anonymizer
| Distro | Build dev | Release |
|---|---|---|
| ROS 2 Humble (u22.04) | ||
| ROS 2 Jazzy (u24.04) | ||
| ROS 2 Kilted (u24.04) | ||
| ROS 2 Lyrical (u26.04) | ||
| ROS 2 Rolling (u26.04) |
Blurs the faces in the camera images of MCAP files (ROS 2 bags), so datasets recorded among people can be shared without privacy issues, while keeping everything else exactly as it was recorded.
-
Only the faces change. In raw images (
sensor_msgs/Image), every pixel outside the blurred faces stays bit-exact, and so do the encoding,step, header and message size. Each encoding is processed in its own format: a Bayer image is still a valid mosaic, a YUV 4:2:2 image keeps its chroma subsampling, 16-bit images keep their depth and byte order. -
Everything else is untouched. All other messages are copied byte by byte, with their log
and publish times and sequence numbers, in the same order. Schemas, channels (also those
without messages, with their QoS metadata), metadata and attachment records, and the chunk
compression are kept. For a rosbag2 directory, its
metadata.yamlis copied. -
Compressed images (
sensor_msgs/CompressedImage): PNG is re-encoded losslessly (same channels and bit depth); JPEG is re-encoded with its own quantization tables and chroma subsampling, so the extra loss in the rest of the image is minimal. - Privacy first. A low detection threshold, boxes enlarged by a margin and propagated to the neighboring frames (so a face missed by the detector in a few frames is still covered), a review video to check the result, and hand-added boxes for anything missed. Topics that may contain images but cannot be processed (e.g. compressed video) stop the tool, instead of being silently copied.
- Large and unfinished bags. Files are streamed (memory use is about one MCAP chunk, whatever the bag size), and recordings that were never closed (no summary section, truncated last chunk) are processed up to their last complete record. The output is always a complete, indexed MCAP file.
Faces are detected with CenterFace at a low threshold (high recall), and doubtful detections are kept only if a person detector (YOLOX-s) finds a person around them, which removes most false positives (foliage, wheels, machinery…). Both are ONNX models that run with OpenCV (no extra dependencies), or with onnxruntime on the CPU or an NVIDIA GPU. All the bundled code and models have permissive licenses (BSD, MIT, Apache-2.0).
Install
ROS 2 (Humble, Jazzy, Kilted, Lyrical, Rolling)
From source, in a colcon workspace:
cd ~/ros2_ws/src
git clone https://github.com/MOLAorg/mcap_anonymizer.git
cd ~/ros2_ws
rosdep install --from-paths src --ignore-src -y
colcon build --packages-select mcap_anonymizer
source install/setup.bash
ros2 run mcap_anonymizer mcap_anonymize --help
Its only dependencies are python3-numpy, python3-opencv and python3-pil; the MCAP reader
and writer (the mcap Python library, MIT) is bundled.
pip (no ROS needed)
pip install 'mcap-anonymizer[compression] @ git+https://github.com/MOLAorg/mcap_anonymizer.git'
mcap_anonymize --help
Optional extras
-
zstd or lz4 compressed MCAP chunks need the Python packages
zstandardorlz4(Ubuntu:sudo apt install python3-zstandard python3-lz4; pip: the[compression]extra). Bags recorded with rosbag2’s default MCAP settings are not compressed. -
Faster detection with onnxruntime:
pip install onnxruntime. It is used automatically when installed (--backend auto). -
NVIDIA GPU:
pip install 'onnxruntime-gpu[cuda,cudnn]'(and notonnxruntime), then--device cuda(orauto). Older GPUs (Pascal: GTX 10xx, Quadro Pxxx) are not supported by the CUDA 13 libraries of the latest versions; on them, usepip install 'onnxruntime-gpu[cuda,cudnn]==1.22.0' 'nvidia-cudnn-cu12==9.5.1.17'.
Quick start
mcap_anonymize run input.mcap output.mcap --detections detections.json
mcap_anonymize run my_rosbag2_dir/ my_rosbag2_dir_anonymized/ # all the .mcap files of a bag
With ROS 2, use ros2 run mcap_anonymizer mcap_anonymize instead of mcap_anonymize.
All the image topics are processed. The tool prints how many images of each topic were
anonymized, and detections.json keeps the detected faces.
Recommended workflow to publish a dataset
```bash
1. Detect the faces (the slow part; see “Performance”)
mcap_anonymize detect input.mcap detections.json
File truncated at 100 lines see the full file
Package Dependencies
System Dependencies
Dependant Packages
Launch files
Messages
Services
Plugins
Recent questions tagged mcap_anonymizer at Robotics Stack Exchange
|
mcap_anonymizer package from mcap_anonymizer repomcap_anonymizer |
ROS Distro
|
Package Summary
| Version | 0.1.0 |
| License | BSD-3-Clause |
| Build type | AMENT_PYTHON |
| Use | RECOMMENDED |
Repository Summary
| Checkout URI | https://github.com/MOLAorg/mcap_anonymizer.git |
| VCS Type | git |
| VCS Version | main |
| Last Updated | 2026-11-18 |
| Dev Status | DEVELOPED |
| Released | UNRELEASED |
| Contributing |
Help Wanted (-)
Good First Issues (-) Pull Requests to Review (-) |
Package Description
Additional Links
Maintainers
- Jose Luis Blanco-Claraco
Authors
- Jose Luis Blanco-Claraco
mcap_anonymizer
| Distro | Build dev | Release |
|---|---|---|
| ROS 2 Humble (u22.04) | ||
| ROS 2 Jazzy (u24.04) | ||
| ROS 2 Kilted (u24.04) | ||
| ROS 2 Lyrical (u26.04) | ||
| ROS 2 Rolling (u26.04) |
Blurs the faces in the camera images of MCAP files (ROS 2 bags), so datasets recorded among people can be shared without privacy issues, while keeping everything else exactly as it was recorded.
-
Only the faces change. In raw images (
sensor_msgs/Image), every pixel outside the blurred faces stays bit-exact, and so do the encoding,step, header and message size. Each encoding is processed in its own format: a Bayer image is still a valid mosaic, a YUV 4:2:2 image keeps its chroma subsampling, 16-bit images keep their depth and byte order. -
Everything else is untouched. All other messages are copied byte by byte, with their log
and publish times and sequence numbers, in the same order. Schemas, channels (also those
without messages, with their QoS metadata), metadata and attachment records, and the chunk
compression are kept. For a rosbag2 directory, its
metadata.yamlis copied. -
Compressed images (
sensor_msgs/CompressedImage): PNG is re-encoded losslessly (same channels and bit depth); JPEG is re-encoded with its own quantization tables and chroma subsampling, so the extra loss in the rest of the image is minimal. - Privacy first. A low detection threshold, boxes enlarged by a margin and propagated to the neighboring frames (so a face missed by the detector in a few frames is still covered), a review video to check the result, and hand-added boxes for anything missed. Topics that may contain images but cannot be processed (e.g. compressed video) stop the tool, instead of being silently copied.
- Large and unfinished bags. Files are streamed (memory use is about one MCAP chunk, whatever the bag size), and recordings that were never closed (no summary section, truncated last chunk) are processed up to their last complete record. The output is always a complete, indexed MCAP file.
Faces are detected with CenterFace at a low threshold (high recall), and doubtful detections are kept only if a person detector (YOLOX-s) finds a person around them, which removes most false positives (foliage, wheels, machinery…). Both are ONNX models that run with OpenCV (no extra dependencies), or with onnxruntime on the CPU or an NVIDIA GPU. All the bundled code and models have permissive licenses (BSD, MIT, Apache-2.0).
Install
ROS 2 (Humble, Jazzy, Kilted, Lyrical, Rolling)
From source, in a colcon workspace:
cd ~/ros2_ws/src
git clone https://github.com/MOLAorg/mcap_anonymizer.git
cd ~/ros2_ws
rosdep install --from-paths src --ignore-src -y
colcon build --packages-select mcap_anonymizer
source install/setup.bash
ros2 run mcap_anonymizer mcap_anonymize --help
Its only dependencies are python3-numpy, python3-opencv and python3-pil; the MCAP reader
and writer (the mcap Python library, MIT) is bundled.
pip (no ROS needed)
pip install 'mcap-anonymizer[compression] @ git+https://github.com/MOLAorg/mcap_anonymizer.git'
mcap_anonymize --help
Optional extras
-
zstd or lz4 compressed MCAP chunks need the Python packages
zstandardorlz4(Ubuntu:sudo apt install python3-zstandard python3-lz4; pip: the[compression]extra). Bags recorded with rosbag2’s default MCAP settings are not compressed. -
Faster detection with onnxruntime:
pip install onnxruntime. It is used automatically when installed (--backend auto). -
NVIDIA GPU:
pip install 'onnxruntime-gpu[cuda,cudnn]'(and notonnxruntime), then--device cuda(orauto). Older GPUs (Pascal: GTX 10xx, Quadro Pxxx) are not supported by the CUDA 13 libraries of the latest versions; on them, usepip install 'onnxruntime-gpu[cuda,cudnn]==1.22.0' 'nvidia-cudnn-cu12==9.5.1.17'.
Quick start
mcap_anonymize run input.mcap output.mcap --detections detections.json
mcap_anonymize run my_rosbag2_dir/ my_rosbag2_dir_anonymized/ # all the .mcap files of a bag
With ROS 2, use ros2 run mcap_anonymizer mcap_anonymize instead of mcap_anonymize.
All the image topics are processed. The tool prints how many images of each topic were
anonymized, and detections.json keeps the detected faces.
Recommended workflow to publish a dataset
```bash
1. Detect the faces (the slow part; see “Performance”)
mcap_anonymize detect input.mcap detections.json
File truncated at 100 lines see the full file
Package Dependencies
System Dependencies
Dependant Packages
Launch files
Messages
Services
Plugins
Recent questions tagged mcap_anonymizer at Robotics Stack Exchange
|
mcap_anonymizer package from mcap_anonymizer repomcap_anonymizer |
ROS Distro
|
Package Summary
| Version | 0.1.0 |
| License | BSD-3-Clause |
| Build type | AMENT_PYTHON |
| Use | RECOMMENDED |
Repository Summary
| Checkout URI | https://github.com/MOLAorg/mcap_anonymizer.git |
| VCS Type | git |
| VCS Version | main |
| Last Updated | 2026-11-18 |
| Dev Status | DEVELOPED |
| Released | UNRELEASED |
| Contributing |
Help Wanted (-)
Good First Issues (-) Pull Requests to Review (-) |
Package Description
Additional Links
Maintainers
- Jose Luis Blanco-Claraco
Authors
- Jose Luis Blanco-Claraco
mcap_anonymizer
| Distro | Build dev | Release |
|---|---|---|
| ROS 2 Humble (u22.04) | ||
| ROS 2 Jazzy (u24.04) | ||
| ROS 2 Kilted (u24.04) | ||
| ROS 2 Lyrical (u26.04) | ||
| ROS 2 Rolling (u26.04) |
Blurs the faces in the camera images of MCAP files (ROS 2 bags), so datasets recorded among people can be shared without privacy issues, while keeping everything else exactly as it was recorded.
-
Only the faces change. In raw images (
sensor_msgs/Image), every pixel outside the blurred faces stays bit-exact, and so do the encoding,step, header and message size. Each encoding is processed in its own format: a Bayer image is still a valid mosaic, a YUV 4:2:2 image keeps its chroma subsampling, 16-bit images keep their depth and byte order. -
Everything else is untouched. All other messages are copied byte by byte, with their log
and publish times and sequence numbers, in the same order. Schemas, channels (also those
without messages, with their QoS metadata), metadata and attachment records, and the chunk
compression are kept. For a rosbag2 directory, its
metadata.yamlis copied. -
Compressed images (
sensor_msgs/CompressedImage): PNG is re-encoded losslessly (same channels and bit depth); JPEG is re-encoded with its own quantization tables and chroma subsampling, so the extra loss in the rest of the image is minimal. - Privacy first. A low detection threshold, boxes enlarged by a margin and propagated to the neighboring frames (so a face missed by the detector in a few frames is still covered), a review video to check the result, and hand-added boxes for anything missed. Topics that may contain images but cannot be processed (e.g. compressed video) stop the tool, instead of being silently copied.
- Large and unfinished bags. Files are streamed (memory use is about one MCAP chunk, whatever the bag size), and recordings that were never closed (no summary section, truncated last chunk) are processed up to their last complete record. The output is always a complete, indexed MCAP file.
Faces are detected with CenterFace at a low threshold (high recall), and doubtful detections are kept only if a person detector (YOLOX-s) finds a person around them, which removes most false positives (foliage, wheels, machinery…). Both are ONNX models that run with OpenCV (no extra dependencies), or with onnxruntime on the CPU or an NVIDIA GPU. All the bundled code and models have permissive licenses (BSD, MIT, Apache-2.0).
Install
ROS 2 (Humble, Jazzy, Kilted, Lyrical, Rolling)
From source, in a colcon workspace:
cd ~/ros2_ws/src
git clone https://github.com/MOLAorg/mcap_anonymizer.git
cd ~/ros2_ws
rosdep install --from-paths src --ignore-src -y
colcon build --packages-select mcap_anonymizer
source install/setup.bash
ros2 run mcap_anonymizer mcap_anonymize --help
Its only dependencies are python3-numpy, python3-opencv and python3-pil; the MCAP reader
and writer (the mcap Python library, MIT) is bundled.
pip (no ROS needed)
pip install 'mcap-anonymizer[compression] @ git+https://github.com/MOLAorg/mcap_anonymizer.git'
mcap_anonymize --help
Optional extras
-
zstd or lz4 compressed MCAP chunks need the Python packages
zstandardorlz4(Ubuntu:sudo apt install python3-zstandard python3-lz4; pip: the[compression]extra). Bags recorded with rosbag2’s default MCAP settings are not compressed. -
Faster detection with onnxruntime:
pip install onnxruntime. It is used automatically when installed (--backend auto). -
NVIDIA GPU:
pip install 'onnxruntime-gpu[cuda,cudnn]'(and notonnxruntime), then--device cuda(orauto). Older GPUs (Pascal: GTX 10xx, Quadro Pxxx) are not supported by the CUDA 13 libraries of the latest versions; on them, usepip install 'onnxruntime-gpu[cuda,cudnn]==1.22.0' 'nvidia-cudnn-cu12==9.5.1.17'.
Quick start
mcap_anonymize run input.mcap output.mcap --detections detections.json
mcap_anonymize run my_rosbag2_dir/ my_rosbag2_dir_anonymized/ # all the .mcap files of a bag
With ROS 2, use ros2 run mcap_anonymizer mcap_anonymize instead of mcap_anonymize.
All the image topics are processed. The tool prints how many images of each topic were
anonymized, and detections.json keeps the detected faces.
Recommended workflow to publish a dataset
```bash
1. Detect the faces (the slow part; see “Performance”)
mcap_anonymize detect input.mcap detections.json
File truncated at 100 lines see the full file
Package Dependencies
System Dependencies
Dependant Packages
Launch files
Messages
Services
Plugins
Recent questions tagged mcap_anonymizer at Robotics Stack Exchange
|
mcap_anonymizer package from mcap_anonymizer repomcap_anonymizer |
ROS Distro
|
Package Summary
| Version | 0.1.0 |
| License | BSD-3-Clause |
| Build type | AMENT_PYTHON |
| Use | RECOMMENDED |
Repository Summary
| Checkout URI | https://github.com/MOLAorg/mcap_anonymizer.git |
| VCS Type | git |
| VCS Version | main |
| Last Updated | 2026-11-18 |
| Dev Status | DEVELOPED |
| Released | UNRELEASED |
| Contributing |
Help Wanted (-)
Good First Issues (-) Pull Requests to Review (-) |
Package Description
Additional Links
Maintainers
- Jose Luis Blanco-Claraco
Authors
- Jose Luis Blanco-Claraco
mcap_anonymizer
| Distro | Build dev | Release |
|---|---|---|
| ROS 2 Humble (u22.04) | ||
| ROS 2 Jazzy (u24.04) | ||
| ROS 2 Kilted (u24.04) | ||
| ROS 2 Lyrical (u26.04) | ||
| ROS 2 Rolling (u26.04) |
Blurs the faces in the camera images of MCAP files (ROS 2 bags), so datasets recorded among people can be shared without privacy issues, while keeping everything else exactly as it was recorded.
-
Only the faces change. In raw images (
sensor_msgs/Image), every pixel outside the blurred faces stays bit-exact, and so do the encoding,step, header and message size. Each encoding is processed in its own format: a Bayer image is still a valid mosaic, a YUV 4:2:2 image keeps its chroma subsampling, 16-bit images keep their depth and byte order. -
Everything else is untouched. All other messages are copied byte by byte, with their log
and publish times and sequence numbers, in the same order. Schemas, channels (also those
without messages, with their QoS metadata), metadata and attachment records, and the chunk
compression are kept. For a rosbag2 directory, its
metadata.yamlis copied. -
Compressed images (
sensor_msgs/CompressedImage): PNG is re-encoded losslessly (same channels and bit depth); JPEG is re-encoded with its own quantization tables and chroma subsampling, so the extra loss in the rest of the image is minimal. - Privacy first. A low detection threshold, boxes enlarged by a margin and propagated to the neighboring frames (so a face missed by the detector in a few frames is still covered), a review video to check the result, and hand-added boxes for anything missed. Topics that may contain images but cannot be processed (e.g. compressed video) stop the tool, instead of being silently copied.
- Large and unfinished bags. Files are streamed (memory use is about one MCAP chunk, whatever the bag size), and recordings that were never closed (no summary section, truncated last chunk) are processed up to their last complete record. The output is always a complete, indexed MCAP file.
Faces are detected with CenterFace at a low threshold (high recall), and doubtful detections are kept only if a person detector (YOLOX-s) finds a person around them, which removes most false positives (foliage, wheels, machinery…). Both are ONNX models that run with OpenCV (no extra dependencies), or with onnxruntime on the CPU or an NVIDIA GPU. All the bundled code and models have permissive licenses (BSD, MIT, Apache-2.0).
Install
ROS 2 (Humble, Jazzy, Kilted, Lyrical, Rolling)
From source, in a colcon workspace:
cd ~/ros2_ws/src
git clone https://github.com/MOLAorg/mcap_anonymizer.git
cd ~/ros2_ws
rosdep install --from-paths src --ignore-src -y
colcon build --packages-select mcap_anonymizer
source install/setup.bash
ros2 run mcap_anonymizer mcap_anonymize --help
Its only dependencies are python3-numpy, python3-opencv and python3-pil; the MCAP reader
and writer (the mcap Python library, MIT) is bundled.
pip (no ROS needed)
pip install 'mcap-anonymizer[compression] @ git+https://github.com/MOLAorg/mcap_anonymizer.git'
mcap_anonymize --help
Optional extras
-
zstd or lz4 compressed MCAP chunks need the Python packages
zstandardorlz4(Ubuntu:sudo apt install python3-zstandard python3-lz4; pip: the[compression]extra). Bags recorded with rosbag2’s default MCAP settings are not compressed. -
Faster detection with onnxruntime:
pip install onnxruntime. It is used automatically when installed (--backend auto). -
NVIDIA GPU:
pip install 'onnxruntime-gpu[cuda,cudnn]'(and notonnxruntime), then--device cuda(orauto). Older GPUs (Pascal: GTX 10xx, Quadro Pxxx) are not supported by the CUDA 13 libraries of the latest versions; on them, usepip install 'onnxruntime-gpu[cuda,cudnn]==1.22.0' 'nvidia-cudnn-cu12==9.5.1.17'.
Quick start
mcap_anonymize run input.mcap output.mcap --detections detections.json
mcap_anonymize run my_rosbag2_dir/ my_rosbag2_dir_anonymized/ # all the .mcap files of a bag
With ROS 2, use ros2 run mcap_anonymizer mcap_anonymize instead of mcap_anonymize.
All the image topics are processed. The tool prints how many images of each topic were
anonymized, and detections.json keeps the detected faces.
Recommended workflow to publish a dataset
```bash
1. Detect the faces (the slow part; see “Performance”)
mcap_anonymize detect input.mcap detections.json
File truncated at 100 lines see the full file
Package Dependencies
System Dependencies
Dependant Packages
Launch files
Messages
Services
Plugins
Recent questions tagged mcap_anonymizer at Robotics Stack Exchange
|
mcap_anonymizer package from mcap_anonymizer repomcap_anonymizer |
ROS Distro
|
Package Summary
| Version | 0.1.0 |
| License | BSD-3-Clause |
| Build type | AMENT_PYTHON |
| Use | RECOMMENDED |
Repository Summary
| Checkout URI | https://github.com/MOLAorg/mcap_anonymizer.git |
| VCS Type | git |
| VCS Version | main |
| Last Updated | 2026-11-18 |
| Dev Status | DEVELOPED |
| Released | UNRELEASED |
| Contributing |
Help Wanted (-)
Good First Issues (-) Pull Requests to Review (-) |
Package Description
Additional Links
Maintainers
- Jose Luis Blanco-Claraco
Authors
- Jose Luis Blanco-Claraco
mcap_anonymizer
| Distro | Build dev | Release |
|---|---|---|
| ROS 2 Humble (u22.04) | ||
| ROS 2 Jazzy (u24.04) | ||
| ROS 2 Kilted (u24.04) | ||
| ROS 2 Lyrical (u26.04) | ||
| ROS 2 Rolling (u26.04) |
Blurs the faces in the camera images of MCAP files (ROS 2 bags), so datasets recorded among people can be shared without privacy issues, while keeping everything else exactly as it was recorded.
-
Only the faces change. In raw images (
sensor_msgs/Image), every pixel outside the blurred faces stays bit-exact, and so do the encoding,step, header and message size. Each encoding is processed in its own format: a Bayer image is still a valid mosaic, a YUV 4:2:2 image keeps its chroma subsampling, 16-bit images keep their depth and byte order. -
Everything else is untouched. All other messages are copied byte by byte, with their log
and publish times and sequence numbers, in the same order. Schemas, channels (also those
without messages, with their QoS metadata), metadata and attachment records, and the chunk
compression are kept. For a rosbag2 directory, its
metadata.yamlis copied. -
Compressed images (
sensor_msgs/CompressedImage): PNG is re-encoded losslessly (same channels and bit depth); JPEG is re-encoded with its own quantization tables and chroma subsampling, so the extra loss in the rest of the image is minimal. - Privacy first. A low detection threshold, boxes enlarged by a margin and propagated to the neighboring frames (so a face missed by the detector in a few frames is still covered), a review video to check the result, and hand-added boxes for anything missed. Topics that may contain images but cannot be processed (e.g. compressed video) stop the tool, instead of being silently copied.
- Large and unfinished bags. Files are streamed (memory use is about one MCAP chunk, whatever the bag size), and recordings that were never closed (no summary section, truncated last chunk) are processed up to their last complete record. The output is always a complete, indexed MCAP file.
Faces are detected with CenterFace at a low threshold (high recall), and doubtful detections are kept only if a person detector (YOLOX-s) finds a person around them, which removes most false positives (foliage, wheels, machinery…). Both are ONNX models that run with OpenCV (no extra dependencies), or with onnxruntime on the CPU or an NVIDIA GPU. All the bundled code and models have permissive licenses (BSD, MIT, Apache-2.0).
Install
ROS 2 (Humble, Jazzy, Kilted, Lyrical, Rolling)
From source, in a colcon workspace:
cd ~/ros2_ws/src
git clone https://github.com/MOLAorg/mcap_anonymizer.git
cd ~/ros2_ws
rosdep install --from-paths src --ignore-src -y
colcon build --packages-select mcap_anonymizer
source install/setup.bash
ros2 run mcap_anonymizer mcap_anonymize --help
Its only dependencies are python3-numpy, python3-opencv and python3-pil; the MCAP reader
and writer (the mcap Python library, MIT) is bundled.
pip (no ROS needed)
pip install 'mcap-anonymizer[compression] @ git+https://github.com/MOLAorg/mcap_anonymizer.git'
mcap_anonymize --help
Optional extras
-
zstd or lz4 compressed MCAP chunks need the Python packages
zstandardorlz4(Ubuntu:sudo apt install python3-zstandard python3-lz4; pip: the[compression]extra). Bags recorded with rosbag2’s default MCAP settings are not compressed. -
Faster detection with onnxruntime:
pip install onnxruntime. It is used automatically when installed (--backend auto). -
NVIDIA GPU:
pip install 'onnxruntime-gpu[cuda,cudnn]'(and notonnxruntime), then--device cuda(orauto). Older GPUs (Pascal: GTX 10xx, Quadro Pxxx) are not supported by the CUDA 13 libraries of the latest versions; on them, usepip install 'onnxruntime-gpu[cuda,cudnn]==1.22.0' 'nvidia-cudnn-cu12==9.5.1.17'.
Quick start
mcap_anonymize run input.mcap output.mcap --detections detections.json
mcap_anonymize run my_rosbag2_dir/ my_rosbag2_dir_anonymized/ # all the .mcap files of a bag
With ROS 2, use ros2 run mcap_anonymizer mcap_anonymize instead of mcap_anonymize.
All the image topics are processed. The tool prints how many images of each topic were
anonymized, and detections.json keeps the detected faces.
Recommended workflow to publish a dataset
```bash
1. Detect the faces (the slow part; see “Performance”)
mcap_anonymize detect input.mcap detections.json
File truncated at 100 lines see the full file
Package Dependencies
System Dependencies
Dependant Packages
Launch files
Messages
Services
Plugins
Recent questions tagged mcap_anonymizer at Robotics Stack Exchange
|
mcap_anonymizer package from mcap_anonymizer repomcap_anonymizer |
ROS Distro
|
Package Summary
| Version | 0.1.0 |
| License | BSD-3-Clause |
| Build type | AMENT_PYTHON |
| Use | RECOMMENDED |
Repository Summary
| Checkout URI | https://github.com/MOLAorg/mcap_anonymizer.git |
| VCS Type | git |
| VCS Version | main |
| Last Updated | 2026-11-18 |
| Dev Status | DEVELOPED |
| Released | UNRELEASED |
| Contributing |
Help Wanted (-)
Good First Issues (-) Pull Requests to Review (-) |
Package Description
Additional Links
Maintainers
- Jose Luis Blanco-Claraco
Authors
- Jose Luis Blanco-Claraco
mcap_anonymizer
| Distro | Build dev | Release |
|---|---|---|
| ROS 2 Humble (u22.04) | ||
| ROS 2 Jazzy (u24.04) | ||
| ROS 2 Kilted (u24.04) | ||
| ROS 2 Lyrical (u26.04) | ||
| ROS 2 Rolling (u26.04) |
Blurs the faces in the camera images of MCAP files (ROS 2 bags), so datasets recorded among people can be shared without privacy issues, while keeping everything else exactly as it was recorded.
-
Only the faces change. In raw images (
sensor_msgs/Image), every pixel outside the blurred faces stays bit-exact, and so do the encoding,step, header and message size. Each encoding is processed in its own format: a Bayer image is still a valid mosaic, a YUV 4:2:2 image keeps its chroma subsampling, 16-bit images keep their depth and byte order. -
Everything else is untouched. All other messages are copied byte by byte, with their log
and publish times and sequence numbers, in the same order. Schemas, channels (also those
without messages, with their QoS metadata), metadata and attachment records, and the chunk
compression are kept. For a rosbag2 directory, its
metadata.yamlis copied. -
Compressed images (
sensor_msgs/CompressedImage): PNG is re-encoded losslessly (same channels and bit depth); JPEG is re-encoded with its own quantization tables and chroma subsampling, so the extra loss in the rest of the image is minimal. - Privacy first. A low detection threshold, boxes enlarged by a margin and propagated to the neighboring frames (so a face missed by the detector in a few frames is still covered), a review video to check the result, and hand-added boxes for anything missed. Topics that may contain images but cannot be processed (e.g. compressed video) stop the tool, instead of being silently copied.
- Large and unfinished bags. Files are streamed (memory use is about one MCAP chunk, whatever the bag size), and recordings that were never closed (no summary section, truncated last chunk) are processed up to their last complete record. The output is always a complete, indexed MCAP file.
Faces are detected with CenterFace at a low threshold (high recall), and doubtful detections are kept only if a person detector (YOLOX-s) finds a person around them, which removes most false positives (foliage, wheels, machinery…). Both are ONNX models that run with OpenCV (no extra dependencies), or with onnxruntime on the CPU or an NVIDIA GPU. All the bundled code and models have permissive licenses (BSD, MIT, Apache-2.0).
Install
ROS 2 (Humble, Jazzy, Kilted, Lyrical, Rolling)
From source, in a colcon workspace:
cd ~/ros2_ws/src
git clone https://github.com/MOLAorg/mcap_anonymizer.git
cd ~/ros2_ws
rosdep install --from-paths src --ignore-src -y
colcon build --packages-select mcap_anonymizer
source install/setup.bash
ros2 run mcap_anonymizer mcap_anonymize --help
Its only dependencies are python3-numpy, python3-opencv and python3-pil; the MCAP reader
and writer (the mcap Python library, MIT) is bundled.
pip (no ROS needed)
pip install 'mcap-anonymizer[compression] @ git+https://github.com/MOLAorg/mcap_anonymizer.git'
mcap_anonymize --help
Optional extras
-
zstd or lz4 compressed MCAP chunks need the Python packages
zstandardorlz4(Ubuntu:sudo apt install python3-zstandard python3-lz4; pip: the[compression]extra). Bags recorded with rosbag2’s default MCAP settings are not compressed. -
Faster detection with onnxruntime:
pip install onnxruntime. It is used automatically when installed (--backend auto). -
NVIDIA GPU:
pip install 'onnxruntime-gpu[cuda,cudnn]'(and notonnxruntime), then--device cuda(orauto). Older GPUs (Pascal: GTX 10xx, Quadro Pxxx) are not supported by the CUDA 13 libraries of the latest versions; on them, usepip install 'onnxruntime-gpu[cuda,cudnn]==1.22.0' 'nvidia-cudnn-cu12==9.5.1.17'.
Quick start
mcap_anonymize run input.mcap output.mcap --detections detections.json
mcap_anonymize run my_rosbag2_dir/ my_rosbag2_dir_anonymized/ # all the .mcap files of a bag
With ROS 2, use ros2 run mcap_anonymizer mcap_anonymize instead of mcap_anonymize.
All the image topics are processed. The tool prints how many images of each topic were
anonymized, and detections.json keeps the detected faces.
Recommended workflow to publish a dataset
```bash
1. Detect the faces (the slow part; see “Performance”)
mcap_anonymize detect input.mcap detections.json
File truncated at 100 lines see the full file
Package Dependencies
System Dependencies
Dependant Packages
Launch files
Messages
Services
Plugins
Recent questions tagged mcap_anonymizer at Robotics Stack Exchange
|
mcap_anonymizer package from mcap_anonymizer repomcap_anonymizer |
ROS Distro
|
Package Summary
| Version | 0.1.0 |
| License | BSD-3-Clause |
| Build type | AMENT_PYTHON |
| Use | RECOMMENDED |
Repository Summary
| Checkout URI | https://github.com/MOLAorg/mcap_anonymizer.git |
| VCS Type | git |
| VCS Version | main |
| Last Updated | 2026-11-18 |
| Dev Status | DEVELOPED |
| Released | UNRELEASED |
| Contributing |
Help Wanted (-)
Good First Issues (-) Pull Requests to Review (-) |
Package Description
Additional Links
Maintainers
- Jose Luis Blanco-Claraco
Authors
- Jose Luis Blanco-Claraco
mcap_anonymizer
| Distro | Build dev | Release |
|---|---|---|
| ROS 2 Humble (u22.04) | ||
| ROS 2 Jazzy (u24.04) | ||
| ROS 2 Kilted (u24.04) | ||
| ROS 2 Lyrical (u26.04) | ||
| ROS 2 Rolling (u26.04) |
Blurs the faces in the camera images of MCAP files (ROS 2 bags), so datasets recorded among people can be shared without privacy issues, while keeping everything else exactly as it was recorded.
-
Only the faces change. In raw images (
sensor_msgs/Image), every pixel outside the blurred faces stays bit-exact, and so do the encoding,step, header and message size. Each encoding is processed in its own format: a Bayer image is still a valid mosaic, a YUV 4:2:2 image keeps its chroma subsampling, 16-bit images keep their depth and byte order. -
Everything else is untouched. All other messages are copied byte by byte, with their log
and publish times and sequence numbers, in the same order. Schemas, channels (also those
without messages, with their QoS metadata), metadata and attachment records, and the chunk
compression are kept. For a rosbag2 directory, its
metadata.yamlis copied. -
Compressed images (
sensor_msgs/CompressedImage): PNG is re-encoded losslessly (same channels and bit depth); JPEG is re-encoded with its own quantization tables and chroma subsampling, so the extra loss in the rest of the image is minimal. - Privacy first. A low detection threshold, boxes enlarged by a margin and propagated to the neighboring frames (so a face missed by the detector in a few frames is still covered), a review video to check the result, and hand-added boxes for anything missed. Topics that may contain images but cannot be processed (e.g. compressed video) stop the tool, instead of being silently copied.
- Large and unfinished bags. Files are streamed (memory use is about one MCAP chunk, whatever the bag size), and recordings that were never closed (no summary section, truncated last chunk) are processed up to their last complete record. The output is always a complete, indexed MCAP file.
Faces are detected with CenterFace at a low threshold (high recall), and doubtful detections are kept only if a person detector (YOLOX-s) finds a person around them, which removes most false positives (foliage, wheels, machinery…). Both are ONNX models that run with OpenCV (no extra dependencies), or with onnxruntime on the CPU or an NVIDIA GPU. All the bundled code and models have permissive licenses (BSD, MIT, Apache-2.0).
Install
ROS 2 (Humble, Jazzy, Kilted, Lyrical, Rolling)
From source, in a colcon workspace:
cd ~/ros2_ws/src
git clone https://github.com/MOLAorg/mcap_anonymizer.git
cd ~/ros2_ws
rosdep install --from-paths src --ignore-src -y
colcon build --packages-select mcap_anonymizer
source install/setup.bash
ros2 run mcap_anonymizer mcap_anonymize --help
Its only dependencies are python3-numpy, python3-opencv and python3-pil; the MCAP reader
and writer (the mcap Python library, MIT) is bundled.
pip (no ROS needed)
pip install 'mcap-anonymizer[compression] @ git+https://github.com/MOLAorg/mcap_anonymizer.git'
mcap_anonymize --help
Optional extras
-
zstd or lz4 compressed MCAP chunks need the Python packages
zstandardorlz4(Ubuntu:sudo apt install python3-zstandard python3-lz4; pip: the[compression]extra). Bags recorded with rosbag2’s default MCAP settings are not compressed. -
Faster detection with onnxruntime:
pip install onnxruntime. It is used automatically when installed (--backend auto). -
NVIDIA GPU:
pip install 'onnxruntime-gpu[cuda,cudnn]'(and notonnxruntime), then--device cuda(orauto). Older GPUs (Pascal: GTX 10xx, Quadro Pxxx) are not supported by the CUDA 13 libraries of the latest versions; on them, usepip install 'onnxruntime-gpu[cuda,cudnn]==1.22.0' 'nvidia-cudnn-cu12==9.5.1.17'.
Quick start
mcap_anonymize run input.mcap output.mcap --detections detections.json
mcap_anonymize run my_rosbag2_dir/ my_rosbag2_dir_anonymized/ # all the .mcap files of a bag
With ROS 2, use ros2 run mcap_anonymizer mcap_anonymize instead of mcap_anonymize.
All the image topics are processed. The tool prints how many images of each topic were
anonymized, and detections.json keeps the detected faces.
Recommended workflow to publish a dataset
```bash
1. Detect the faces (the slow part; see “Performance”)
mcap_anonymize detect input.mcap detections.json
File truncated at 100 lines see the full file
Package Dependencies
System Dependencies
Dependant Packages
Launch files
Messages
Services
Plugins
Recent questions tagged mcap_anonymizer at Robotics Stack Exchange
|
mcap_anonymizer package from mcap_anonymizer repomcap_anonymizer |
ROS Distro
|
Package Summary
| Version | 0.1.0 |
| License | BSD-3-Clause |
| Build type | AMENT_PYTHON |
| Use | RECOMMENDED |
Repository Summary
| Checkout URI | https://github.com/MOLAorg/mcap_anonymizer.git |
| VCS Type | git |
| VCS Version | main |
| Last Updated | 2026-11-18 |
| Dev Status | DEVELOPED |
| Released | UNRELEASED |
| Contributing |
Help Wanted (-)
Good First Issues (-) Pull Requests to Review (-) |
Package Description
Additional Links
Maintainers
- Jose Luis Blanco-Claraco
Authors
- Jose Luis Blanco-Claraco
mcap_anonymizer
| Distro | Build dev | Release |
|---|---|---|
| ROS 2 Humble (u22.04) | ||
| ROS 2 Jazzy (u24.04) | ||
| ROS 2 Kilted (u24.04) | ||
| ROS 2 Lyrical (u26.04) | ||
| ROS 2 Rolling (u26.04) |
Blurs the faces in the camera images of MCAP files (ROS 2 bags), so datasets recorded among people can be shared without privacy issues, while keeping everything else exactly as it was recorded.
-
Only the faces change. In raw images (
sensor_msgs/Image), every pixel outside the blurred faces stays bit-exact, and so do the encoding,step, header and message size. Each encoding is processed in its own format: a Bayer image is still a valid mosaic, a YUV 4:2:2 image keeps its chroma subsampling, 16-bit images keep their depth and byte order. -
Everything else is untouched. All other messages are copied byte by byte, with their log
and publish times and sequence numbers, in the same order. Schemas, channels (also those
without messages, with their QoS metadata), metadata and attachment records, and the chunk
compression are kept. For a rosbag2 directory, its
metadata.yamlis copied. -
Compressed images (
sensor_msgs/CompressedImage): PNG is re-encoded losslessly (same channels and bit depth); JPEG is re-encoded with its own quantization tables and chroma subsampling, so the extra loss in the rest of the image is minimal. - Privacy first. A low detection threshold, boxes enlarged by a margin and propagated to the neighboring frames (so a face missed by the detector in a few frames is still covered), a review video to check the result, and hand-added boxes for anything missed. Topics that may contain images but cannot be processed (e.g. compressed video) stop the tool, instead of being silently copied.
- Large and unfinished bags. Files are streamed (memory use is about one MCAP chunk, whatever the bag size), and recordings that were never closed (no summary section, truncated last chunk) are processed up to their last complete record. The output is always a complete, indexed MCAP file.
Faces are detected with CenterFace at a low threshold (high recall), and doubtful detections are kept only if a person detector (YOLOX-s) finds a person around them, which removes most false positives (foliage, wheels, machinery…). Both are ONNX models that run with OpenCV (no extra dependencies), or with onnxruntime on the CPU or an NVIDIA GPU. All the bundled code and models have permissive licenses (BSD, MIT, Apache-2.0).
Install
ROS 2 (Humble, Jazzy, Kilted, Lyrical, Rolling)
From source, in a colcon workspace:
cd ~/ros2_ws/src
git clone https://github.com/MOLAorg/mcap_anonymizer.git
cd ~/ros2_ws
rosdep install --from-paths src --ignore-src -y
colcon build --packages-select mcap_anonymizer
source install/setup.bash
ros2 run mcap_anonymizer mcap_anonymize --help
Its only dependencies are python3-numpy, python3-opencv and python3-pil; the MCAP reader
and writer (the mcap Python library, MIT) is bundled.
pip (no ROS needed)
pip install 'mcap-anonymizer[compression] @ git+https://github.com/MOLAorg/mcap_anonymizer.git'
mcap_anonymize --help
Optional extras
-
zstd or lz4 compressed MCAP chunks need the Python packages
zstandardorlz4(Ubuntu:sudo apt install python3-zstandard python3-lz4; pip: the[compression]extra). Bags recorded with rosbag2’s default MCAP settings are not compressed. -
Faster detection with onnxruntime:
pip install onnxruntime. It is used automatically when installed (--backend auto). -
NVIDIA GPU:
pip install 'onnxruntime-gpu[cuda,cudnn]'(and notonnxruntime), then--device cuda(orauto). Older GPUs (Pascal: GTX 10xx, Quadro Pxxx) are not supported by the CUDA 13 libraries of the latest versions; on them, usepip install 'onnxruntime-gpu[cuda,cudnn]==1.22.0' 'nvidia-cudnn-cu12==9.5.1.17'.
Quick start
mcap_anonymize run input.mcap output.mcap --detections detections.json
mcap_anonymize run my_rosbag2_dir/ my_rosbag2_dir_anonymized/ # all the .mcap files of a bag
With ROS 2, use ros2 run mcap_anonymizer mcap_anonymize instead of mcap_anonymize.
All the image topics are processed. The tool prints how many images of each topic were
anonymized, and detections.json keeps the detected faces.
Recommended workflow to publish a dataset
```bash
1. Detect the faces (the slow part; see “Performance”)
mcap_anonymize detect input.mcap detections.json
File truncated at 100 lines see the full file
Package Dependencies
System Dependencies
Dependant Packages
Launch files
Messages
Services
Plugins
Recent questions tagged mcap_anonymizer at Robotics Stack Exchange
|
mcap_anonymizer package from mcap_anonymizer repomcap_anonymizer |
ROS Distro
|
Package Summary
| Version | 0.1.0 |
| License | BSD-3-Clause |
| Build type | AMENT_PYTHON |
| Use | RECOMMENDED |
Repository Summary
| Checkout URI | https://github.com/MOLAorg/mcap_anonymizer.git |
| VCS Type | git |
| VCS Version | main |
| Last Updated | 2026-11-18 |
| Dev Status | DEVELOPED |
| Released | UNRELEASED |
| Contributing |
Help Wanted (-)
Good First Issues (-) Pull Requests to Review (-) |
Package Description
Additional Links
Maintainers
- Jose Luis Blanco-Claraco
Authors
- Jose Luis Blanco-Claraco
mcap_anonymizer
| Distro | Build dev | Release |
|---|---|---|
| ROS 2 Humble (u22.04) | ||
| ROS 2 Jazzy (u24.04) | ||
| ROS 2 Kilted (u24.04) | ||
| ROS 2 Lyrical (u26.04) | ||
| ROS 2 Rolling (u26.04) |
Blurs the faces in the camera images of MCAP files (ROS 2 bags), so datasets recorded among people can be shared without privacy issues, while keeping everything else exactly as it was recorded.
-
Only the faces change. In raw images (
sensor_msgs/Image), every pixel outside the blurred faces stays bit-exact, and so do the encoding,step, header and message size. Each encoding is processed in its own format: a Bayer image is still a valid mosaic, a YUV 4:2:2 image keeps its chroma subsampling, 16-bit images keep their depth and byte order. -
Everything else is untouched. All other messages are copied byte by byte, with their log
and publish times and sequence numbers, in the same order. Schemas, channels (also those
without messages, with their QoS metadata), metadata and attachment records, and the chunk
compression are kept. For a rosbag2 directory, its
metadata.yamlis copied. -
Compressed images (
sensor_msgs/CompressedImage): PNG is re-encoded losslessly (same channels and bit depth); JPEG is re-encoded with its own quantization tables and chroma subsampling, so the extra loss in the rest of the image is minimal. - Privacy first. A low detection threshold, boxes enlarged by a margin and propagated to the neighboring frames (so a face missed by the detector in a few frames is still covered), a review video to check the result, and hand-added boxes for anything missed. Topics that may contain images but cannot be processed (e.g. compressed video) stop the tool, instead of being silently copied.
- Large and unfinished bags. Files are streamed (memory use is about one MCAP chunk, whatever the bag size), and recordings that were never closed (no summary section, truncated last chunk) are processed up to their last complete record. The output is always a complete, indexed MCAP file.
Faces are detected with CenterFace at a low threshold (high recall), and doubtful detections are kept only if a person detector (YOLOX-s) finds a person around them, which removes most false positives (foliage, wheels, machinery…). Both are ONNX models that run with OpenCV (no extra dependencies), or with onnxruntime on the CPU or an NVIDIA GPU. All the bundled code and models have permissive licenses (BSD, MIT, Apache-2.0).
Install
ROS 2 (Humble, Jazzy, Kilted, Lyrical, Rolling)
From source, in a colcon workspace:
cd ~/ros2_ws/src
git clone https://github.com/MOLAorg/mcap_anonymizer.git
cd ~/ros2_ws
rosdep install --from-paths src --ignore-src -y
colcon build --packages-select mcap_anonymizer
source install/setup.bash
ros2 run mcap_anonymizer mcap_anonymize --help
Its only dependencies are python3-numpy, python3-opencv and python3-pil; the MCAP reader
and writer (the mcap Python library, MIT) is bundled.
pip (no ROS needed)
pip install 'mcap-anonymizer[compression] @ git+https://github.com/MOLAorg/mcap_anonymizer.git'
mcap_anonymize --help
Optional extras
-
zstd or lz4 compressed MCAP chunks need the Python packages
zstandardorlz4(Ubuntu:sudo apt install python3-zstandard python3-lz4; pip: the[compression]extra). Bags recorded with rosbag2’s default MCAP settings are not compressed. -
Faster detection with onnxruntime:
pip install onnxruntime. It is used automatically when installed (--backend auto). -
NVIDIA GPU:
pip install 'onnxruntime-gpu[cuda,cudnn]'(and notonnxruntime), then--device cuda(orauto). Older GPUs (Pascal: GTX 10xx, Quadro Pxxx) are not supported by the CUDA 13 libraries of the latest versions; on them, usepip install 'onnxruntime-gpu[cuda,cudnn]==1.22.0' 'nvidia-cudnn-cu12==9.5.1.17'.
Quick start
mcap_anonymize run input.mcap output.mcap --detections detections.json
mcap_anonymize run my_rosbag2_dir/ my_rosbag2_dir_anonymized/ # all the .mcap files of a bag
With ROS 2, use ros2 run mcap_anonymizer mcap_anonymize instead of mcap_anonymize.
All the image topics are processed. The tool prints how many images of each topic were
anonymized, and detections.json keeps the detected faces.
Recommended workflow to publish a dataset
```bash
1. Detect the faces (the slow part; see “Performance”)
mcap_anonymize detect input.mcap detections.json
File truncated at 100 lines see the full file
Package Dependencies
System Dependencies
Dependant Packages
Launch files
Messages
Services
Plugins
Recent questions tagged mcap_anonymizer at Robotics Stack Exchange
|
mcap_anonymizer package from mcap_anonymizer repomcap_anonymizer |
ROS Distro
|
Package Summary
| Version | 0.1.0 |
| License | BSD-3-Clause |
| Build type | AMENT_PYTHON |
| Use | RECOMMENDED |
Repository Summary
| Checkout URI | https://github.com/MOLAorg/mcap_anonymizer.git |
| VCS Type | git |
| VCS Version | main |
| Last Updated | 2026-11-18 |
| Dev Status | DEVELOPED |
| Released | UNRELEASED |
| Contributing |
Help Wanted (-)
Good First Issues (-) Pull Requests to Review (-) |
Package Description
Additional Links
Maintainers
- Jose Luis Blanco-Claraco
Authors
- Jose Luis Blanco-Claraco
mcap_anonymizer
| Distro | Build dev | Release |
|---|---|---|
| ROS 2 Humble (u22.04) | ||
| ROS 2 Jazzy (u24.04) | ||
| ROS 2 Kilted (u24.04) | ||
| ROS 2 Lyrical (u26.04) | ||
| ROS 2 Rolling (u26.04) |
Blurs the faces in the camera images of MCAP files (ROS 2 bags), so datasets recorded among people can be shared without privacy issues, while keeping everything else exactly as it was recorded.
-
Only the faces change. In raw images (
sensor_msgs/Image), every pixel outside the blurred faces stays bit-exact, and so do the encoding,step, header and message size. Each encoding is processed in its own format: a Bayer image is still a valid mosaic, a YUV 4:2:2 image keeps its chroma subsampling, 16-bit images keep their depth and byte order. -
Everything else is untouched. All other messages are copied byte by byte, with their log
and publish times and sequence numbers, in the same order. Schemas, channels (also those
without messages, with their QoS metadata), metadata and attachment records, and the chunk
compression are kept. For a rosbag2 directory, its
metadata.yamlis copied. -
Compressed images (
sensor_msgs/CompressedImage): PNG is re-encoded losslessly (same channels and bit depth); JPEG is re-encoded with its own quantization tables and chroma subsampling, so the extra loss in the rest of the image is minimal. - Privacy first. A low detection threshold, boxes enlarged by a margin and propagated to the neighboring frames (so a face missed by the detector in a few frames is still covered), a review video to check the result, and hand-added boxes for anything missed. Topics that may contain images but cannot be processed (e.g. compressed video) stop the tool, instead of being silently copied.
- Large and unfinished bags. Files are streamed (memory use is about one MCAP chunk, whatever the bag size), and recordings that were never closed (no summary section, truncated last chunk) are processed up to their last complete record. The output is always a complete, indexed MCAP file.
Faces are detected with CenterFace at a low threshold (high recall), and doubtful detections are kept only if a person detector (YOLOX-s) finds a person around them, which removes most false positives (foliage, wheels, machinery…). Both are ONNX models that run with OpenCV (no extra dependencies), or with onnxruntime on the CPU or an NVIDIA GPU. All the bundled code and models have permissive licenses (BSD, MIT, Apache-2.0).
Install
ROS 2 (Humble, Jazzy, Kilted, Lyrical, Rolling)
From source, in a colcon workspace:
cd ~/ros2_ws/src
git clone https://github.com/MOLAorg/mcap_anonymizer.git
cd ~/ros2_ws
rosdep install --from-paths src --ignore-src -y
colcon build --packages-select mcap_anonymizer
source install/setup.bash
ros2 run mcap_anonymizer mcap_anonymize --help
Its only dependencies are python3-numpy, python3-opencv and python3-pil; the MCAP reader
and writer (the mcap Python library, MIT) is bundled.
pip (no ROS needed)
pip install 'mcap-anonymizer[compression] @ git+https://github.com/MOLAorg/mcap_anonymizer.git'
mcap_anonymize --help
Optional extras
-
zstd or lz4 compressed MCAP chunks need the Python packages
zstandardorlz4(Ubuntu:sudo apt install python3-zstandard python3-lz4; pip: the[compression]extra). Bags recorded with rosbag2’s default MCAP settings are not compressed. -
Faster detection with onnxruntime:
pip install onnxruntime. It is used automatically when installed (--backend auto). -
NVIDIA GPU:
pip install 'onnxruntime-gpu[cuda,cudnn]'(and notonnxruntime), then--device cuda(orauto). Older GPUs (Pascal: GTX 10xx, Quadro Pxxx) are not supported by the CUDA 13 libraries of the latest versions; on them, usepip install 'onnxruntime-gpu[cuda,cudnn]==1.22.0' 'nvidia-cudnn-cu12==9.5.1.17'.
Quick start
mcap_anonymize run input.mcap output.mcap --detections detections.json
mcap_anonymize run my_rosbag2_dir/ my_rosbag2_dir_anonymized/ # all the .mcap files of a bag
With ROS 2, use ros2 run mcap_anonymizer mcap_anonymize instead of mcap_anonymize.
All the image topics are processed. The tool prints how many images of each topic were
anonymized, and detections.json keeps the detected faces.
Recommended workflow to publish a dataset
```bash
1. Detect the faces (the slow part; see “Performance”)
mcap_anonymize detect input.mcap detections.json
File truncated at 100 lines see the full file
Package Dependencies
System Dependencies
Dependant Packages
Launch files
Messages
Services
Plugins
Recent questions tagged mcap_anonymizer at Robotics Stack Exchange
|
mcap_anonymizer package from mcap_anonymizer repomcap_anonymizer |
ROS Distro
|
Package Summary
| Version | 0.1.0 |
| License | BSD-3-Clause |
| Build type | AMENT_PYTHON |
| Use | RECOMMENDED |
Repository Summary
| Checkout URI | https://github.com/MOLAorg/mcap_anonymizer.git |
| VCS Type | git |
| VCS Version | main |
| Last Updated | 2026-11-18 |
| Dev Status | DEVELOPED |
| Released | UNRELEASED |
| Contributing |
Help Wanted (-)
Good First Issues (-) Pull Requests to Review (-) |
Package Description
Additional Links
Maintainers
- Jose Luis Blanco-Claraco
Authors
- Jose Luis Blanco-Claraco
mcap_anonymizer
| Distro | Build dev | Release |
|---|---|---|
| ROS 2 Humble (u22.04) | ||
| ROS 2 Jazzy (u24.04) | ||
| ROS 2 Kilted (u24.04) | ||
| ROS 2 Lyrical (u26.04) | ||
| ROS 2 Rolling (u26.04) |
Blurs the faces in the camera images of MCAP files (ROS 2 bags), so datasets recorded among people can be shared without privacy issues, while keeping everything else exactly as it was recorded.
-
Only the faces change. In raw images (
sensor_msgs/Image), every pixel outside the blurred faces stays bit-exact, and so do the encoding,step, header and message size. Each encoding is processed in its own format: a Bayer image is still a valid mosaic, a YUV 4:2:2 image keeps its chroma subsampling, 16-bit images keep their depth and byte order. -
Everything else is untouched. All other messages are copied byte by byte, with their log
and publish times and sequence numbers, in the same order. Schemas, channels (also those
without messages, with their QoS metadata), metadata and attachment records, and the chunk
compression are kept. For a rosbag2 directory, its
metadata.yamlis copied. -
Compressed images (
sensor_msgs/CompressedImage): PNG is re-encoded losslessly (same channels and bit depth); JPEG is re-encoded with its own quantization tables and chroma subsampling, so the extra loss in the rest of the image is minimal. - Privacy first. A low detection threshold, boxes enlarged by a margin and propagated to the neighboring frames (so a face missed by the detector in a few frames is still covered), a review video to check the result, and hand-added boxes for anything missed. Topics that may contain images but cannot be processed (e.g. compressed video) stop the tool, instead of being silently copied.
- Large and unfinished bags. Files are streamed (memory use is about one MCAP chunk, whatever the bag size), and recordings that were never closed (no summary section, truncated last chunk) are processed up to their last complete record. The output is always a complete, indexed MCAP file.
Faces are detected with CenterFace at a low threshold (high recall), and doubtful detections are kept only if a person detector (YOLOX-s) finds a person around them, which removes most false positives (foliage, wheels, machinery…). Both are ONNX models that run with OpenCV (no extra dependencies), or with onnxruntime on the CPU or an NVIDIA GPU. All the bundled code and models have permissive licenses (BSD, MIT, Apache-2.0).
Install
ROS 2 (Humble, Jazzy, Kilted, Lyrical, Rolling)
From source, in a colcon workspace:
cd ~/ros2_ws/src
git clone https://github.com/MOLAorg/mcap_anonymizer.git
cd ~/ros2_ws
rosdep install --from-paths src --ignore-src -y
colcon build --packages-select mcap_anonymizer
source install/setup.bash
ros2 run mcap_anonymizer mcap_anonymize --help
Its only dependencies are python3-numpy, python3-opencv and python3-pil; the MCAP reader
and writer (the mcap Python library, MIT) is bundled.
pip (no ROS needed)
pip install 'mcap-anonymizer[compression] @ git+https://github.com/MOLAorg/mcap_anonymizer.git'
mcap_anonymize --help
Optional extras
-
zstd or lz4 compressed MCAP chunks need the Python packages
zstandardorlz4(Ubuntu:sudo apt install python3-zstandard python3-lz4; pip: the[compression]extra). Bags recorded with rosbag2’s default MCAP settings are not compressed. -
Faster detection with onnxruntime:
pip install onnxruntime. It is used automatically when installed (--backend auto). -
NVIDIA GPU:
pip install 'onnxruntime-gpu[cuda,cudnn]'(and notonnxruntime), then--device cuda(orauto). Older GPUs (Pascal: GTX 10xx, Quadro Pxxx) are not supported by the CUDA 13 libraries of the latest versions; on them, usepip install 'onnxruntime-gpu[cuda,cudnn]==1.22.0' 'nvidia-cudnn-cu12==9.5.1.17'.
Quick start
mcap_anonymize run input.mcap output.mcap --detections detections.json
mcap_anonymize run my_rosbag2_dir/ my_rosbag2_dir_anonymized/ # all the .mcap files of a bag
With ROS 2, use ros2 run mcap_anonymizer mcap_anonymize instead of mcap_anonymize.
All the image topics are processed. The tool prints how many images of each topic were
anonymized, and detections.json keeps the detected faces.
Recommended workflow to publish a dataset
```bash
1. Detect the faces (the slow part; see “Performance”)
mcap_anonymize detect input.mcap detections.json
File truncated at 100 lines see the full file
Package Dependencies
System Dependencies
Dependant Packages
Launch files
Messages
Services
Plugins
Recent questions tagged mcap_anonymizer at Robotics Stack Exchange
|
mcap_anonymizer package from mcap_anonymizer repomcap_anonymizer |
ROS Distro
|
Package Summary
| Version | 0.1.0 |
| License | BSD-3-Clause |
| Build type | AMENT_PYTHON |
| Use | RECOMMENDED |
Repository Summary
| Checkout URI | https://github.com/MOLAorg/mcap_anonymizer.git |
| VCS Type | git |
| VCS Version | main |
| Last Updated | 2026-11-18 |
| Dev Status | DEVELOPED |
| Released | UNRELEASED |
| Contributing |
Help Wanted (-)
Good First Issues (-) Pull Requests to Review (-) |
Package Description
Additional Links
Maintainers
- Jose Luis Blanco-Claraco
Authors
- Jose Luis Blanco-Claraco
mcap_anonymizer
| Distro | Build dev | Release |
|---|---|---|
| ROS 2 Humble (u22.04) | ||
| ROS 2 Jazzy (u24.04) | ||
| ROS 2 Kilted (u24.04) | ||
| ROS 2 Lyrical (u26.04) | ||
| ROS 2 Rolling (u26.04) |
Blurs the faces in the camera images of MCAP files (ROS 2 bags), so datasets recorded among people can be shared without privacy issues, while keeping everything else exactly as it was recorded.
-
Only the faces change. In raw images (
sensor_msgs/Image), every pixel outside the blurred faces stays bit-exact, and so do the encoding,step, header and message size. Each encoding is processed in its own format: a Bayer image is still a valid mosaic, a YUV 4:2:2 image keeps its chroma subsampling, 16-bit images keep their depth and byte order. -
Everything else is untouched. All other messages are copied byte by byte, with their log
and publish times and sequence numbers, in the same order. Schemas, channels (also those
without messages, with their QoS metadata), metadata and attachment records, and the chunk
compression are kept. For a rosbag2 directory, its
metadata.yamlis copied. -
Compressed images (
sensor_msgs/CompressedImage): PNG is re-encoded losslessly (same channels and bit depth); JPEG is re-encoded with its own quantization tables and chroma subsampling, so the extra loss in the rest of the image is minimal. - Privacy first. A low detection threshold, boxes enlarged by a margin and propagated to the neighboring frames (so a face missed by the detector in a few frames is still covered), a review video to check the result, and hand-added boxes for anything missed. Topics that may contain images but cannot be processed (e.g. compressed video) stop the tool, instead of being silently copied.
- Large and unfinished bags. Files are streamed (memory use is about one MCAP chunk, whatever the bag size), and recordings that were never closed (no summary section, truncated last chunk) are processed up to their last complete record. The output is always a complete, indexed MCAP file.
Faces are detected with CenterFace at a low threshold (high recall), and doubtful detections are kept only if a person detector (YOLOX-s) finds a person around them, which removes most false positives (foliage, wheels, machinery…). Both are ONNX models that run with OpenCV (no extra dependencies), or with onnxruntime on the CPU or an NVIDIA GPU. All the bundled code and models have permissive licenses (BSD, MIT, Apache-2.0).
Install
ROS 2 (Humble, Jazzy, Kilted, Lyrical, Rolling)
From source, in a colcon workspace:
cd ~/ros2_ws/src
git clone https://github.com/MOLAorg/mcap_anonymizer.git
cd ~/ros2_ws
rosdep install --from-paths src --ignore-src -y
colcon build --packages-select mcap_anonymizer
source install/setup.bash
ros2 run mcap_anonymizer mcap_anonymize --help
Its only dependencies are python3-numpy, python3-opencv and python3-pil; the MCAP reader
and writer (the mcap Python library, MIT) is bundled.
pip (no ROS needed)
pip install 'mcap-anonymizer[compression] @ git+https://github.com/MOLAorg/mcap_anonymizer.git'
mcap_anonymize --help
Optional extras
-
zstd or lz4 compressed MCAP chunks need the Python packages
zstandardorlz4(Ubuntu:sudo apt install python3-zstandard python3-lz4; pip: the[compression]extra). Bags recorded with rosbag2’s default MCAP settings are not compressed. -
Faster detection with onnxruntime:
pip install onnxruntime. It is used automatically when installed (--backend auto). -
NVIDIA GPU:
pip install 'onnxruntime-gpu[cuda,cudnn]'(and notonnxruntime), then--device cuda(orauto). Older GPUs (Pascal: GTX 10xx, Quadro Pxxx) are not supported by the CUDA 13 libraries of the latest versions; on them, usepip install 'onnxruntime-gpu[cuda,cudnn]==1.22.0' 'nvidia-cudnn-cu12==9.5.1.17'.
Quick start
mcap_anonymize run input.mcap output.mcap --detections detections.json
mcap_anonymize run my_rosbag2_dir/ my_rosbag2_dir_anonymized/ # all the .mcap files of a bag
With ROS 2, use ros2 run mcap_anonymizer mcap_anonymize instead of mcap_anonymize.
All the image topics are processed. The tool prints how many images of each topic were
anonymized, and detections.json keeps the detected faces.
Recommended workflow to publish a dataset
```bash
1. Detect the faces (the slow part; see “Performance”)
mcap_anonymize detect input.mcap detections.json
File truncated at 100 lines see the full file
Package Dependencies
System Dependencies
Dependant Packages
Launch files
Messages
Services
Plugins
Recent questions tagged mcap_anonymizer at Robotics Stack Exchange
|
mcap_anonymizer package from mcap_anonymizer repomcap_anonymizer |
ROS Distro
|
Package Summary
| Version | 0.1.0 |
| License | BSD-3-Clause |
| Build type | AMENT_PYTHON |
| Use | RECOMMENDED |
Repository Summary
| Checkout URI | https://github.com/MOLAorg/mcap_anonymizer.git |
| VCS Type | git |
| VCS Version | main |
| Last Updated | 2026-11-18 |
| Dev Status | DEVELOPED |
| Released | UNRELEASED |
| Contributing |
Help Wanted (-)
Good First Issues (-) Pull Requests to Review (-) |
Package Description
Additional Links
Maintainers
- Jose Luis Blanco-Claraco
Authors
- Jose Luis Blanco-Claraco
mcap_anonymizer
| Distro | Build dev | Release |
|---|---|---|
| ROS 2 Humble (u22.04) | ||
| ROS 2 Jazzy (u24.04) | ||
| ROS 2 Kilted (u24.04) | ||
| ROS 2 Lyrical (u26.04) | ||
| ROS 2 Rolling (u26.04) |
Blurs the faces in the camera images of MCAP files (ROS 2 bags), so datasets recorded among people can be shared without privacy issues, while keeping everything else exactly as it was recorded.
-
Only the faces change. In raw images (
sensor_msgs/Image), every pixel outside the blurred faces stays bit-exact, and so do the encoding,step, header and message size. Each encoding is processed in its own format: a Bayer image is still a valid mosaic, a YUV 4:2:2 image keeps its chroma subsampling, 16-bit images keep their depth and byte order. -
Everything else is untouched. All other messages are copied byte by byte, with their log
and publish times and sequence numbers, in the same order. Schemas, channels (also those
without messages, with their QoS metadata), metadata and attachment records, and the chunk
compression are kept. For a rosbag2 directory, its
metadata.yamlis copied. -
Compressed images (
sensor_msgs/CompressedImage): PNG is re-encoded losslessly (same channels and bit depth); JPEG is re-encoded with its own quantization tables and chroma subsampling, so the extra loss in the rest of the image is minimal. - Privacy first. A low detection threshold, boxes enlarged by a margin and propagated to the neighboring frames (so a face missed by the detector in a few frames is still covered), a review video to check the result, and hand-added boxes for anything missed. Topics that may contain images but cannot be processed (e.g. compressed video) stop the tool, instead of being silently copied.
- Large and unfinished bags. Files are streamed (memory use is about one MCAP chunk, whatever the bag size), and recordings that were never closed (no summary section, truncated last chunk) are processed up to their last complete record. The output is always a complete, indexed MCAP file.
Faces are detected with CenterFace at a low threshold (high recall), and doubtful detections are kept only if a person detector (YOLOX-s) finds a person around them, which removes most false positives (foliage, wheels, machinery…). Both are ONNX models that run with OpenCV (no extra dependencies), or with onnxruntime on the CPU or an NVIDIA GPU. All the bundled code and models have permissive licenses (BSD, MIT, Apache-2.0).
Install
ROS 2 (Humble, Jazzy, Kilted, Lyrical, Rolling)
From source, in a colcon workspace:
cd ~/ros2_ws/src
git clone https://github.com/MOLAorg/mcap_anonymizer.git
cd ~/ros2_ws
rosdep install --from-paths src --ignore-src -y
colcon build --packages-select mcap_anonymizer
source install/setup.bash
ros2 run mcap_anonymizer mcap_anonymize --help
Its only dependencies are python3-numpy, python3-opencv and python3-pil; the MCAP reader
and writer (the mcap Python library, MIT) is bundled.
pip (no ROS needed)
pip install 'mcap-anonymizer[compression] @ git+https://github.com/MOLAorg/mcap_anonymizer.git'
mcap_anonymize --help
Optional extras
-
zstd or lz4 compressed MCAP chunks need the Python packages
zstandardorlz4(Ubuntu:sudo apt install python3-zstandard python3-lz4; pip: the[compression]extra). Bags recorded with rosbag2’s default MCAP settings are not compressed. -
Faster detection with onnxruntime:
pip install onnxruntime. It is used automatically when installed (--backend auto). -
NVIDIA GPU:
pip install 'onnxruntime-gpu[cuda,cudnn]'(and notonnxruntime), then--device cuda(orauto). Older GPUs (Pascal: GTX 10xx, Quadro Pxxx) are not supported by the CUDA 13 libraries of the latest versions; on them, usepip install 'onnxruntime-gpu[cuda,cudnn]==1.22.0' 'nvidia-cudnn-cu12==9.5.1.17'.
Quick start
mcap_anonymize run input.mcap output.mcap --detections detections.json
mcap_anonymize run my_rosbag2_dir/ my_rosbag2_dir_anonymized/ # all the .mcap files of a bag
With ROS 2, use ros2 run mcap_anonymizer mcap_anonymize instead of mcap_anonymize.
All the image topics are processed. The tool prints how many images of each topic were
anonymized, and detections.json keeps the detected faces.
Recommended workflow to publish a dataset
```bash
1. Detect the faces (the slow part; see “Performance”)
mcap_anonymize detect input.mcap detections.json
File truncated at 100 lines see the full file
Package Dependencies
System Dependencies
Dependant Packages
Launch files
Messages
Services
Plugins
Recent questions tagged mcap_anonymizer at Robotics Stack Exchange
|
mcap_anonymizer package from mcap_anonymizer repomcap_anonymizer |
ROS Distro
|
Package Summary
| Version | 0.1.0 |
| License | BSD-3-Clause |
| Build type | AMENT_PYTHON |
| Use | RECOMMENDED |
Repository Summary
| Checkout URI | https://github.com/MOLAorg/mcap_anonymizer.git |
| VCS Type | git |
| VCS Version | main |
| Last Updated | 2026-11-18 |
| Dev Status | DEVELOPED |
| Released | UNRELEASED |
| Contributing |
Help Wanted (-)
Good First Issues (-) Pull Requests to Review (-) |
Package Description
Additional Links
Maintainers
- Jose Luis Blanco-Claraco
Authors
- Jose Luis Blanco-Claraco
mcap_anonymizer
| Distro | Build dev | Release |
|---|---|---|
| ROS 2 Humble (u22.04) | ||
| ROS 2 Jazzy (u24.04) | ||
| ROS 2 Kilted (u24.04) | ||
| ROS 2 Lyrical (u26.04) | ||
| ROS 2 Rolling (u26.04) |
Blurs the faces in the camera images of MCAP files (ROS 2 bags), so datasets recorded among people can be shared without privacy issues, while keeping everything else exactly as it was recorded.
-
Only the faces change. In raw images (
sensor_msgs/Image), every pixel outside the blurred faces stays bit-exact, and so do the encoding,step, header and message size. Each encoding is processed in its own format: a Bayer image is still a valid mosaic, a YUV 4:2:2 image keeps its chroma subsampling, 16-bit images keep their depth and byte order. -
Everything else is untouched. All other messages are copied byte by byte, with their log
and publish times and sequence numbers, in the same order. Schemas, channels (also those
without messages, with their QoS metadata), metadata and attachment records, and the chunk
compression are kept. For a rosbag2 directory, its
metadata.yamlis copied. -
Compressed images (
sensor_msgs/CompressedImage): PNG is re-encoded losslessly (same channels and bit depth); JPEG is re-encoded with its own quantization tables and chroma subsampling, so the extra loss in the rest of the image is minimal. - Privacy first. A low detection threshold, boxes enlarged by a margin and propagated to the neighboring frames (so a face missed by the detector in a few frames is still covered), a review video to check the result, and hand-added boxes for anything missed. Topics that may contain images but cannot be processed (e.g. compressed video) stop the tool, instead of being silently copied.
- Large and unfinished bags. Files are streamed (memory use is about one MCAP chunk, whatever the bag size), and recordings that were never closed (no summary section, truncated last chunk) are processed up to their last complete record. The output is always a complete, indexed MCAP file.
Faces are detected with CenterFace at a low threshold (high recall), and doubtful detections are kept only if a person detector (YOLOX-s) finds a person around them, which removes most false positives (foliage, wheels, machinery…). Both are ONNX models that run with OpenCV (no extra dependencies), or with onnxruntime on the CPU or an NVIDIA GPU. All the bundled code and models have permissive licenses (BSD, MIT, Apache-2.0).
Install
ROS 2 (Humble, Jazzy, Kilted, Lyrical, Rolling)
From source, in a colcon workspace:
cd ~/ros2_ws/src
git clone https://github.com/MOLAorg/mcap_anonymizer.git
cd ~/ros2_ws
rosdep install --from-paths src --ignore-src -y
colcon build --packages-select mcap_anonymizer
source install/setup.bash
ros2 run mcap_anonymizer mcap_anonymize --help
Its only dependencies are python3-numpy, python3-opencv and python3-pil; the MCAP reader
and writer (the mcap Python library, MIT) is bundled.
pip (no ROS needed)
pip install 'mcap-anonymizer[compression] @ git+https://github.com/MOLAorg/mcap_anonymizer.git'
mcap_anonymize --help
Optional extras
-
zstd or lz4 compressed MCAP chunks need the Python packages
zstandardorlz4(Ubuntu:sudo apt install python3-zstandard python3-lz4; pip: the[compression]extra). Bags recorded with rosbag2’s default MCAP settings are not compressed. -
Faster detection with onnxruntime:
pip install onnxruntime. It is used automatically when installed (--backend auto). -
NVIDIA GPU:
pip install 'onnxruntime-gpu[cuda,cudnn]'(and notonnxruntime), then--device cuda(orauto). Older GPUs (Pascal: GTX 10xx, Quadro Pxxx) are not supported by the CUDA 13 libraries of the latest versions; on them, usepip install 'onnxruntime-gpu[cuda,cudnn]==1.22.0' 'nvidia-cudnn-cu12==9.5.1.17'.
Quick start
mcap_anonymize run input.mcap output.mcap --detections detections.json
mcap_anonymize run my_rosbag2_dir/ my_rosbag2_dir_anonymized/ # all the .mcap files of a bag
With ROS 2, use ros2 run mcap_anonymizer mcap_anonymize instead of mcap_anonymize.
All the image topics are processed. The tool prints how many images of each topic were
anonymized, and detections.json keeps the detected faces.
Recommended workflow to publish a dataset
```bash
1. Detect the faces (the slow part; see “Performance”)
mcap_anonymize detect input.mcap detections.json
File truncated at 100 lines see the full file
Package Dependencies
System Dependencies
Dependant Packages
Launch files
Messages
Services
Plugins
Recent questions tagged mcap_anonymizer at Robotics Stack Exchange
|
mcap_anonymizer package from mcap_anonymizer repomcap_anonymizer |
ROS Distro
|
Package Summary
| Version | 0.1.0 |
| License | BSD-3-Clause |
| Build type | AMENT_PYTHON |
| Use | RECOMMENDED |
Repository Summary
| Checkout URI | https://github.com/MOLAorg/mcap_anonymizer.git |
| VCS Type | git |
| VCS Version | main |
| Last Updated | 2026-11-18 |
| Dev Status | DEVELOPED |
| Released | UNRELEASED |
| Contributing |
Help Wanted (-)
Good First Issues (-) Pull Requests to Review (-) |
Package Description
Additional Links
Maintainers
- Jose Luis Blanco-Claraco
Authors
- Jose Luis Blanco-Claraco
mcap_anonymizer
| Distro | Build dev | Release |
|---|---|---|
| ROS 2 Humble (u22.04) | ||
| ROS 2 Jazzy (u24.04) | ||
| ROS 2 Kilted (u24.04) | ||
| ROS 2 Lyrical (u26.04) | ||
| ROS 2 Rolling (u26.04) |
Blurs the faces in the camera images of MCAP files (ROS 2 bags), so datasets recorded among people can be shared without privacy issues, while keeping everything else exactly as it was recorded.
-
Only the faces change. In raw images (
sensor_msgs/Image), every pixel outside the blurred faces stays bit-exact, and so do the encoding,step, header and message size. Each encoding is processed in its own format: a Bayer image is still a valid mosaic, a YUV 4:2:2 image keeps its chroma subsampling, 16-bit images keep their depth and byte order. -
Everything else is untouched. All other messages are copied byte by byte, with their log
and publish times and sequence numbers, in the same order. Schemas, channels (also those
without messages, with their QoS metadata), metadata and attachment records, and the chunk
compression are kept. For a rosbag2 directory, its
metadata.yamlis copied. -
Compressed images (
sensor_msgs/CompressedImage): PNG is re-encoded losslessly (same channels and bit depth); JPEG is re-encoded with its own quantization tables and chroma subsampling, so the extra loss in the rest of the image is minimal. - Privacy first. A low detection threshold, boxes enlarged by a margin and propagated to the neighboring frames (so a face missed by the detector in a few frames is still covered), a review video to check the result, and hand-added boxes for anything missed. Topics that may contain images but cannot be processed (e.g. compressed video) stop the tool, instead of being silently copied.
- Large and unfinished bags. Files are streamed (memory use is about one MCAP chunk, whatever the bag size), and recordings that were never closed (no summary section, truncated last chunk) are processed up to their last complete record. The output is always a complete, indexed MCAP file.
Faces are detected with CenterFace at a low threshold (high recall), and doubtful detections are kept only if a person detector (YOLOX-s) finds a person around them, which removes most false positives (foliage, wheels, machinery…). Both are ONNX models that run with OpenCV (no extra dependencies), or with onnxruntime on the CPU or an NVIDIA GPU. All the bundled code and models have permissive licenses (BSD, MIT, Apache-2.0).
Install
ROS 2 (Humble, Jazzy, Kilted, Lyrical, Rolling)
From source, in a colcon workspace:
cd ~/ros2_ws/src
git clone https://github.com/MOLAorg/mcap_anonymizer.git
cd ~/ros2_ws
rosdep install --from-paths src --ignore-src -y
colcon build --packages-select mcap_anonymizer
source install/setup.bash
ros2 run mcap_anonymizer mcap_anonymize --help
Its only dependencies are python3-numpy, python3-opencv and python3-pil; the MCAP reader
and writer (the mcap Python library, MIT) is bundled.
pip (no ROS needed)
pip install 'mcap-anonymizer[compression] @ git+https://github.com/MOLAorg/mcap_anonymizer.git'
mcap_anonymize --help
Optional extras
-
zstd or lz4 compressed MCAP chunks need the Python packages
zstandardorlz4(Ubuntu:sudo apt install python3-zstandard python3-lz4; pip: the[compression]extra). Bags recorded with rosbag2’s default MCAP settings are not compressed. -
Faster detection with onnxruntime:
pip install onnxruntime. It is used automatically when installed (--backend auto). -
NVIDIA GPU:
pip install 'onnxruntime-gpu[cuda,cudnn]'(and notonnxruntime), then--device cuda(orauto). Older GPUs (Pascal: GTX 10xx, Quadro Pxxx) are not supported by the CUDA 13 libraries of the latest versions; on them, usepip install 'onnxruntime-gpu[cuda,cudnn]==1.22.0' 'nvidia-cudnn-cu12==9.5.1.17'.
Quick start
mcap_anonymize run input.mcap output.mcap --detections detections.json
mcap_anonymize run my_rosbag2_dir/ my_rosbag2_dir_anonymized/ # all the .mcap files of a bag
With ROS 2, use ros2 run mcap_anonymizer mcap_anonymize instead of mcap_anonymize.
All the image topics are processed. The tool prints how many images of each topic were
anonymized, and detections.json keeps the detected faces.
Recommended workflow to publish a dataset
```bash
1. Detect the faces (the slow part; see “Performance”)
mcap_anonymize detect input.mcap detections.json
File truncated at 100 lines see the full file
Package Dependencies
System Dependencies
Dependant Packages
Launch files
Messages
Services
Plugins
Recent questions tagged mcap_anonymizer at Robotics Stack Exchange
|
mcap_anonymizer package from mcap_anonymizer repomcap_anonymizer |
ROS Distro
|
Package Summary
| Version | 0.1.0 |
| License | BSD-3-Clause |
| Build type | AMENT_PYTHON |
| Use | RECOMMENDED |
Repository Summary
| Checkout URI | https://github.com/MOLAorg/mcap_anonymizer.git |
| VCS Type | git |
| VCS Version | main |
| Last Updated | 2026-11-18 |
| Dev Status | DEVELOPED |
| Released | UNRELEASED |
| Contributing |
Help Wanted (-)
Good First Issues (-) Pull Requests to Review (-) |
Package Description
Additional Links
Maintainers
- Jose Luis Blanco-Claraco
Authors
- Jose Luis Blanco-Claraco
mcap_anonymizer
| Distro | Build dev | Release |
|---|---|---|
| ROS 2 Humble (u22.04) | ||
| ROS 2 Jazzy (u24.04) | ||
| ROS 2 Kilted (u24.04) | ||
| ROS 2 Lyrical (u26.04) | ||
| ROS 2 Rolling (u26.04) |
Blurs the faces in the camera images of MCAP files (ROS 2 bags), so datasets recorded among people can be shared without privacy issues, while keeping everything else exactly as it was recorded.
-
Only the faces change. In raw images (
sensor_msgs/Image), every pixel outside the blurred faces stays bit-exact, and so do the encoding,step, header and message size. Each encoding is processed in its own format: a Bayer image is still a valid mosaic, a YUV 4:2:2 image keeps its chroma subsampling, 16-bit images keep their depth and byte order. -
Everything else is untouched. All other messages are copied byte by byte, with their log
and publish times and sequence numbers, in the same order. Schemas, channels (also those
without messages, with their QoS metadata), metadata and attachment records, and the chunk
compression are kept. For a rosbag2 directory, its
metadata.yamlis copied. -
Compressed images (
sensor_msgs/CompressedImage): PNG is re-encoded losslessly (same channels and bit depth); JPEG is re-encoded with its own quantization tables and chroma subsampling, so the extra loss in the rest of the image is minimal. - Privacy first. A low detection threshold, boxes enlarged by a margin and propagated to the neighboring frames (so a face missed by the detector in a few frames is still covered), a review video to check the result, and hand-added boxes for anything missed. Topics that may contain images but cannot be processed (e.g. compressed video) stop the tool, instead of being silently copied.
- Large and unfinished bags. Files are streamed (memory use is about one MCAP chunk, whatever the bag size), and recordings that were never closed (no summary section, truncated last chunk) are processed up to their last complete record. The output is always a complete, indexed MCAP file.
Faces are detected with CenterFace at a low threshold (high recall), and doubtful detections are kept only if a person detector (YOLOX-s) finds a person around them, which removes most false positives (foliage, wheels, machinery…). Both are ONNX models that run with OpenCV (no extra dependencies), or with onnxruntime on the CPU or an NVIDIA GPU. All the bundled code and models have permissive licenses (BSD, MIT, Apache-2.0).
Install
ROS 2 (Humble, Jazzy, Kilted, Lyrical, Rolling)
From source, in a colcon workspace:
cd ~/ros2_ws/src
git clone https://github.com/MOLAorg/mcap_anonymizer.git
cd ~/ros2_ws
rosdep install --from-paths src --ignore-src -y
colcon build --packages-select mcap_anonymizer
source install/setup.bash
ros2 run mcap_anonymizer mcap_anonymize --help
Its only dependencies are python3-numpy, python3-opencv and python3-pil; the MCAP reader
and writer (the mcap Python library, MIT) is bundled.
pip (no ROS needed)
pip install 'mcap-anonymizer[compression] @ git+https://github.com/MOLAorg/mcap_anonymizer.git'
mcap_anonymize --help
Optional extras
-
zstd or lz4 compressed MCAP chunks need the Python packages
zstandardorlz4(Ubuntu:sudo apt install python3-zstandard python3-lz4; pip: the[compression]extra). Bags recorded with rosbag2’s default MCAP settings are not compressed. -
Faster detection with onnxruntime:
pip install onnxruntime. It is used automatically when installed (--backend auto). -
NVIDIA GPU:
pip install 'onnxruntime-gpu[cuda,cudnn]'(and notonnxruntime), then--device cuda(orauto). Older GPUs (Pascal: GTX 10xx, Quadro Pxxx) are not supported by the CUDA 13 libraries of the latest versions; on them, usepip install 'onnxruntime-gpu[cuda,cudnn]==1.22.0' 'nvidia-cudnn-cu12==9.5.1.17'.
Quick start
mcap_anonymize run input.mcap output.mcap --detections detections.json
mcap_anonymize run my_rosbag2_dir/ my_rosbag2_dir_anonymized/ # all the .mcap files of a bag
With ROS 2, use ros2 run mcap_anonymizer mcap_anonymize instead of mcap_anonymize.
All the image topics are processed. The tool prints how many images of each topic were
anonymized, and detections.json keeps the detected faces.
Recommended workflow to publish a dataset
```bash
1. Detect the faces (the slow part; see “Performance”)
mcap_anonymize detect input.mcap detections.json
File truncated at 100 lines see the full file