| name | run-ego-reconstruction-video |
| description | Run the egocentric reconstruction pipeline on an input video. Use when a user asks Codex to process an MP4 or video with the reconstruction ego pipeline, choose between HaMeR/WiLoR and DynHaMR hand tracking, pass an object prompt or object mesh, enable undistortion, SLAM, gravity alignment, gsplat refinement, or export/check the final result bundle and Three.js scene. |
Run Ego Reconstruction Video
Decide the Run Mode
Use the consolidated entrypoint unless the user explicitly asks for a legacy script:
cd reconstruction
python modules/v2d_pipelines/run_ego_reconstruction.py --help
Default to --hand_tracking hamer for new work. It uses WiLoR/SAM2/HaMeR,
supports a caller-provided object mesh, and supports gsplat refinement. Use
--hand_tracking dynhamr only when the user needs the legacy ViPE + DynHaMR path
and the manual MANO/BMC assets are present.
Before running, gather or infer:
- input video path, preferably absolute or relative to
reconstruction/
- object prompt, such as
"a tissue box"
- output directory
- whether the video is fisheye/wide angle and should use
--undistort
- whether to use
--object_mesh for an existing metric OBJ
- whether to enable
--run_droid_slam, --run_gravity_alignment, --run_gsplat_refinement, or --export_threejs_result
If setup may be incomplete, use $ego-reconstruction-setup first.
Standard Commands
Minimal HaMeR/WiLoR run:
cd reconstruction
python modules/v2d_pipelines/run_ego_reconstruction.py \
--video data/path/to/video.mp4 \
--output_dir data/outputs/my_run \
--object_prompt \
--hand_tracking hamer \
--reference_frame 0
