Stereo and 3D

Stereo takes (take0002_L.mp4 + take0002_R.mp4) are self-contained: frame pairing, motion data and the camera's factory calibration are read from the files themselves.

Depth video

python3 scripts/stereo_depth_video.py captures/take0002 out_depth.mp4 --wls --imu --ema 0.5

Renders the rectified left and right images with a metric depth map (semi-global block matching; --wls smoothing needs opencv-contrib-python), optionally with a motion strip.

Motion HUD

python3 scripts/stereo_motion_hud.py captures/take0002 out_hud.mp4 --start-s 10 --end-s 25

Side-by-side playback with a reference grid, live rotation rates and the predicted rolling-shutter shear for each frame — useful for judging motion blur and skew (and for seeing why Stereo GS avoids it).

Stereo-inertial odometry (OpenVINS)

docker build -t trinet-openvins:latest scripts/openvins-docker/
python3 scripts/make_openvins_config.py captures/take0002_L.mp4 WORK/ov_config
./scripts/run_openvins.sh WORK
python3 scripts/plot_trajectory.py WORK/ov_out/traj_est.txt

make_openvins_config.py turns the recording's embedded calibration into the configuration files OpenVINS expects, so a recording carries everything a VIO run needs. See the script's help for preparing the input bag.

More

The repository also contains experimental scripts for learned stereo depth, map fusion and Gaussian splatting (stereo_depth_neural.py, stereo_map_fusion.py, gs_*.py). Run any script with --help for its options.

For your own pipelines, the reference calibrations and their Kalibr export work with VINS-Fusion, Basalt and similar tools.