Run Ego Reconstruction Video

SkillMedia

Run this repository’s egocentric reconstruction pipeline on a video. Use when a user asks to process an MP4 with HaMeR or DynHaMR hand tracking, an object prompt or mesh, undistortion, DROID-SLAM, gravity alignment, gsplat refinement, or a Three.js result export.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Run Ego Reconstruction Video skill

What this skill tells your AI

The instructions your AI receives, as published by nvidia-isaac/video_to_data in .claude/skills/run-ego-reconstruction-video/SKILL.md and read by ahel’s review.

Run from reconstruction/. Prefer the consolidated entrypoint and use HaMeR unless the user explicitly needs legacy DynHaMR:

python modules/v2d_pipelines/run_ego_reconstruction.py \
  --video <video.mp4> \
  --output_dir data/outputs/<run> \
  --object_prompt "<object>" \
  --hand_tracking hamer \
  --reference_frame 0

Add --object_mesh <mesh.obj> --skip_object_scale_estimation for a supplied mesh. Add --undistort, --run_droid_slam, --run_gravity_alignment, --run_gsplat_refinement, and --export_threejs_result only when requested.

Re-run the same command to resume cached stages. Confirm result/result.npz and result/mesh.obj; with post-processing, use the final suffixed result directory. Use the setup skill first when images, weights, Docker, or GPU prerequisites are missing.

Signals

GitHub stars
587
Forks
57
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
run-ego-reconstruction-video
Source
github.com/nvidia-isaac/video_to_data