VR-VLA/VR-egodex-inpaint-cache-full

EgoDex Inpaint Cache

Per-frame hand segmentation masks and hand-erased (inpainted) frames for the EgoDex LeRobot v2.1 conversion. These are the visual-alignment inputs for robot retargeting: removing the human hands leaves a clean background plate a rendered robot arm can be composited onto.

Source video: bi199797/VR-egodex-apple-lerobot.

Contents

	episodes	tasks	size
train/	2,575	129	469 GB
test/	909	111	207 GB
total	3,484	240	~678 GB

Episodes average 315 frames (10.5 s at 30 fps).

This is a subset, not the whole corpus. The source has 318,082 episodes; 3,484 of them (1.1%) are inpainted here. Inpainting costs roughly 10 GPU-minutes per episode, so the full set is not practical to produce — see Generation below.

Task coverage

All 240 source tasks are represented, but depth per task is very uneven:

• Most tasks hold a handful of episodes. Work was ordered round-robin by episode index across every task, so partial runs spread evenly rather than exhausting a few tasks.
• train/add_remove_lid_v2.1 is far deeper than everything else (166 GB on its own), from a separate depth-first pass.

Check per-task contents rather than assuming uniform depth.

Format

One tar archive per task, not loose files:

train/<task>.tar
test/<task>.tar

Each archive expands to:

<split>/<task>/masks/episode<N>/frame%04d.png                 # hand segmentation mask
<split>/<task>/inpaint_diffueraser/episode<N>/frame%04d.png   # hands erased

Episode numbering matches the source dataset's episode_%06d.mp4. The <split> prefix is part of the path on purpose: 212 task names appear in both train/ and test/ with overlapping episode numbers.

Archives are uncompressed (tar, no gzip) — the payload is PNG, so compression buys almost nothing and costs CPU on both ends.

Download and extract one task

hf download VR-VLA/VR-egodex-inpaint-cache-full train/clean_surface_v2.1.tar \
    --repo-type dataset --local-dir .
tar -xf train/clean_surface_v2.1.tar

The one split archive

train/add_remove_lid_v2.1.tar is 166 GB, above the Hub's 50 GB per-file limit, so it is stored as five .part files. Reassemble before extracting:

hf download VR-VLA/VR-egodex-inpaint-cache-full --repo-type dataset \
    --include "train/add_remove_lid_v2.1.tar.*.part" --local-dir .
cat train/add_remove_lid_v2.1.tar.*.part > train/add_remove_lid_v2.1.tar
tar -xf train/add_remove_lid_v2.1.tar

Why archives

Uploading this as ~1.2 million loose PNGs was bound by a fixed per-commit cost rather than by bandwidth, and did not converge: it moved ~2 GB in 7 hours. Packed as one archive per task the same data transferred in about 25 minutes. The trade-off is that individual frames can no longer be previewed on the Hub.

Generation

stage	backend	share of runtime
segmentation	SAM2 (per-frame)	~25%
inpainting	DiffuEraser	~75%

Measured on H200s: ~5 episodes/hour per GPU, about 10-12 GPU-minutes per episode — roughly 75x slower than real time, since a ~10 s clip costs ~10 GPU-minutes across both stages. Recent runs: 1,051 episodes in 34.5 h on 6 GPUs, then 623 in 28 h on 4 GPUs.

DiffuEraser is used over ProPainter because ProPainter leaves visible ghosted hand residue on frames with sustained occlusion, at ~2-2.5x lower per-frame cost.

Caveats

• Episodes shorter than 23 frames have masks but no inpainted output — DiffuEraser fails below that length (effective video duration is too short). ~0.6% of episodes.
• Masks and inpainted frames are independent stages; an episode can have masks/ without inpaint_diffueraser/.
• No depth stage in this cache.