Skip to main content

run-ego-reconstruction-video

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.

Quellinformationen

Repository
nvidia-isaac/video_to_data
Letzte Quellaktivität
14. August 2026 um 19:50
Erkannte Sprache von SKILL.md
Englisch
Sterne
847
Forks
92

Installationsoptionen

Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.

Quelldateien prüfen

Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.

Datei-Explorer
3 Dateien

SKILL.md wird angezeigt

SKILL.md
Quellanweisungen · Schreibgeschützte Vorschau
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: ```bash 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: ```bash cd reconstruction python modules/v2d_pipelines/run_ego_reconstruction.py \ --video data/path/to/video.mp4 \ --output_dir data/outputs/my_run \ --object_prompt "a handled object" \ --hand_tracking hamer \ --reference_frame 0 ``` Fuller HaMeR/WiLoR run with common postprocessing: ```bash cd reconstruction python modules/v2d_pipelines/run_ego_reconstruction.py \ --video data/path/to/video.mp4 \ --output_dir data/outputs/my_run \ --object_prompt "a handled object" \ --hand_tracking hamer \ --reference_frame 0 \ --undistort \ --run_droid_slam \ --run_gravity_alignment \ --run_gsplat_refinement \ --export_threejs_result ``` HaMeR/WiLoR with a provided object mesh: ```bash cd reconstruction python modules/v2d_pipelines/run_ego_reconstruction.py \ --video data/path/to/video.mp4 \ --output_dir data/outputs/my_mesh_run \ --object_prompt "a handled object" \ --object_mesh /absolute/path/to/object.obj \ --skip_object_scale_estimation \ --hand_tracking hamer ``` DynHaMR path: ```bash cd reconstruction python modules/v2d_pipelines/run_ego_reconstruction.py \ --video data/path/to/video.mp4 \ --output_dir data/outputs/my_dynhamr_run \ --object_prompt "a handled object" \ --hand_tracking dynhamr \ --undistort ``` Do not combine `--hand_tracking dynhamr` with `--object_mesh` or `--run_gsplat_refinement`; the consolidated runner rejects those combinations. ## Reruns and Partial Outputs Rerun the exact same command after an interruption. The runner checks many stage outputs and prints `[skip]` for completed stages. Use `--dev` only when actively editing module source and wanting containers to mount local code. For repo examples, inspect `reconstruction/data/garage/tissue_box/run_wilor_pipeline.sh`. Use it as a shape reference, not as a hard-coded path template. ## Verify Results Expect the main portable bundle at one of these directories: - `<output_dir>/result/` - `<output_dir>/result_slam/` when DROID-SLAM is enabled - `<output_dir>/result_gravity_aligned/` or `<output_dir>/result_slam_gravity_aligned/` when gravity alignment is enabled A complete result bundle contains at least: ```text result.npz mesh.obj ``` If `--export_threejs_result` was used, inspect: ```text <final_result_dir>/threejs_scene/index.html ``` Useful quick checks: ```bash test -s data/outputs/my_run/result/result.npz test -s data/outputs/my_run/result/mesh.obj ls data/outputs/my_run/*overlay*.mp4 ``` ## Debugging - Missing Python module: re-run `bash reconstruction/scripts/install_ego_reconstruction_packages.sh` in the active environment. - Missing Docker image: re-run `bash reconstruction/scripts/build_ego_reconstruction_packages.sh`. - Missing weights: use `$ego-reconstruction-setup` and choose the download mode that matches the run mode. - No object detection: adjust `--object_prompt` and/or `--reference_frame`. - Bad fisheye geometry: add `--undistort` and re-run into a fresh output directory. - Bad final motion or background camera: try `--run_droid_slam`; add `--run_gravity_alignment` when world-up matters. - Refinement failures: first produce a non-refined `result/`, then add `--run_gsplat_refinement` after the base run is healthy. See `references/run-ego-reconstruction-video.md` for entrypoint details and output map.
Auf GitHub ansehen