mono_depth_viewer

Qt viewer for monocular depth estimation on image topics recorded in ROS 2 bags (sqlite3 .db3). Runs transformer models (Depth Anything V1, DPT-Hybrid MiDaS) from Hugging Face and shows the camera image next to a colourised depth map.

Install (ROS 2 Humble)

cd ~/ros2_ws/src
git clone <this-repo-url> mono_depth_viewer
cd ~/ros2_ws
rosdep install --from-paths src --ignore-src -y
pip install -r src/mono_depth_viewer/requirements.txt
colcon build --packages-select mono_depth_viewer
source install/setup.bash

Use

ros2 run mono_depth_viewer viewer /path/to/bag_folder
# or
ros2 launch mono_depth_viewer viewer.launch.py bag:=/path/to/bag_folder

Pick the topic (default /camera/image_raw), model, input size and colormap in the toolbar. You can play, scrub, save a frame (PNG + raw .npy) or export a side-by-side MP4. The first run downloads model weights from Hugging Face.

Notes

• Output is relative inverse depth (larger = closer), not metric distance.
• On CPU the Small model runs at roughly 1 FPS at 364 px; a CUDA GPU is much faster.
• Only .db3 bags are supported (no .mcap).
• Model licenses: Depth Anything V1 Small is Apache-2.0; Base/Large are CC-BY-NC-4.0 (non-commercial); DPT-Hybrid is MIT. This package’s code is Apache-2.0.

Test

colcon test --packages-select mono_depth_viewer && colcon test-result --verbose

Changelog for package mono_depth_viewer

0.1.1 (2026-10-23)

• Initial release.
• Qt viewer that reads sensor_msgs/Image and CompressedImage topics from ROS 2 sqlite3 (.db3) bags.
• Monocular depth estimation with transformer models (Depth Anything V1 Small/Base/Large, DPT-Hybrid MiDaS).
• Playback, frame scrubbing, frame skipping, colormap and input size selection.
• Save frames (PNG + raw .npy) and export side-by-side MP4 videos.
• Contributors: Divagar