annotate_3d_tracking

Working repository for a long-form, real-world 3D point-tracking dataset built from Xperience-10M: which episodes are used, the moving-object masks, and (later) the GPT-6 annotation requests and answers. This is not a public release.

Access and terms

Everything here is derived from Xperience-10M, a gated dataset released for research and non-commercial use. Access to this repository is reviewed manually and given only to collaborators who are also approved for Xperience-10M. Do not redistribute these files or upload them anywhere else. No raw video or depth is stored here: download it from ropedia-ai/xperience-10m (revision ce943cf) with your own approved account.

Layout

path	what
lists/episodes.csv	one row per episode: category, frames at 10 fps, hours, raw-data path, note
lists/base_A_built.txt	761 episodes, 31.7 h: masks, SLAM, moved-verdict threshold < 1; episode files already built
lists/base_B_threshold_ge1.txt	573 episodes, 27.1 h: masks and SLAM, threshold >= 1 (usable, see the labelling rule)
lists/base_D_needs_gmos.txt / .jsonl	1,176 episodes, 94.8 h: SLAM, usable depth, no pose jumps; masks still to be made. The .jsonl is in motion-seg's list format (file sizes and SHA-256 at revision ce943cf)
masks/xperience/<episode>/	moving-object masks of 1,669 episodes (instances.npz, decisions.json, cameras.npz, manifest.json, provenance.json)
masks/xperience/README.md	format of those files; known_issues.json and provenance.csv alongside
requests/<batch>/, predictions/<batch>/	later: GPT-6 requests (requests.jsonl + images/, the benchmark format) and the answers

Episode names are <capture uuid>_ep<n>; the raw files are <capture uuid>/ep<n>/annotation.hdf5 and stereo_left.mp4 in Xperience-10M. Mask frame t is frame 2t of stereo_left.mp4 (10 fps, 512x512); the depth is the same view at 256x256. Rows marked excluded in episodes.csv (no SLAM, unusable depth, no instances, too short, or SLAM jumps above 5 m/s) are not used.

Labelling rule

• Static: only pixels outside every GMOS instance, and only where several views agree (multi-view consistency). A GMOS instance never joins the static set, whatever its verdict; anything uncertain stays out. Static points are accumulated over the whole episode.
• Dynamic: every GMOS instance. Its points are used per frame (the current frame only), never accumulated.
• The moved / static verdicts in decisions.json are kept for statistics only. In episodes with threshold >= 1 they cannot report "moved": the clip's background noise floor (stereo depth errors on large textureless surfaces) is above the largest possible statistic.

Making the masks for base D (Rivanna)

The code is motion-seg/ in vlm_3d_tracking_annotator (setup, checkpoints and the Slurm jobs are in its README). From motion-seg/:

huggingface-cli download 2inf/annotate_3d_tracking --repo-type dataset --include "lists/*" --local-dir hf_lists
env/bin/python download_data.py --root /scratch/$USER/xperience-10m --list hf_lists/lists/base_D_needs_gmos.jsonl
DATA=/scratch/$USER/xperience-10m OUT=/scratch/$USER/motion_seg_out \
  EXTRA="--list hf_lists/lists/base_D_needs_gmos.jsonl" sbatch --array=0-1175%32 --mem=96G --time=08:00:00 slurm/run.sbatch

These episodes are about 5 minutes long (most of them 4.8-5.0 min), longer than the earlier runs. motion-seg's own estimate for a 5-minute clip: 76 GB peak host RAM (hence --mem=96G), 2.8-5.0 h on an A100-80GB, 1.2-2.2 h on a B200 (about 2,000 B200 hours for all 1,176). Upload each finished episode folder as masks/xperience/<episode>/ in the layout above.

Download

huggingface-cli download 2inf/annotate_3d_tracking --repo-type dataset --include "lists/*" --local-dir annotate_3d_tracking
huggingface-cli download 2inf/annotate_3d_tracking --repo-type dataset --include "masks/xperience/<episode>/*" --local-dir annotate_3d_tracking