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/	5,246	129	~1.0 TB
test/	1,417	111	~310 GB
total	6,663	240	~1.3 TB

~19.2 hours of inpainted video (2,078,255 frames at 30 fps). Episodes average 312 frames (~10.4 s).

This is a subset of the source corpus, which holds 318,082 episodes (~928 h). 6,663 of them (2.1%) are inpainted here.

Task coverage

All 240 source tasks are represented. Depth per task is uneven:

• 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 deeper than the rest (178 GB on its own), from a separate depth-first pass.
• A subset of episodes has been processed twice, independently, for quality comparison — see Format below.

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 (pass 1)
<split>/<task>/inpaint/episode<N>/frame%04d.png              # hands erased (pass 2, subset of episodes)

Episode numbering matches the source dataset's episode_%06d.mp4. The <split> prefix is part of the path because 212 task names appear in both train/ and test/ with overlapping episode numbers. An episode may have masks without any inpainted output, or without both inpaint passes — the two passes are independent and not every episode has been run through both.

Archives are uncompressed (tar, no gzip); the payload is PNG.

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 largest archive

train/add_remove_lid_v2.1.tar is 178 GB from a separate depth-first pass. It downloads like any other archive:

hf download VR-VLA/VR-egodex-inpaint-cache-full train/add_remove_lid_v2.1.tar \
    --repo-type dataset --local-dir .

Older train/add_remove_lid_v2.1.tar.*.part files from a previous split upload are still present in the repo. They are a stale, earlier snapshot — prefer the single .tar above.

Generation

Each episode goes through hand segmentation, then hand removal (inpainting) using the segmented mask. A subset of episodes has been run through two independent inpainting passes for quality comparison, hence the two inpaint*/ directories in Format above.

Processing cost

Recent measurement, doubled-up worker configuration:

throughput	~130-235 episodes/hour (6-12 GPUs)
per episode	~2.7-3.2 GPU-minutes

Actual cost varies with episode length and which processing pass(es) an episode goes through.

Why archives

The cache is ~2 million loose PNGs. Uploading it file-by-file is bound by a fixed per-commit cost rather than by bandwidth. Packed as one archive per task the same data transfers in well under an hour. Individual frames cannot be previewed on the Hub as a result.

Caveats

• Very short episodes have masks but no inpainted output — inpainting needs a minimum clip length to run. A small fraction of episodes fall below it.
• Masks and inpainted frames are independent stages; an episode can have masks/ without either inpaint/ or inpaint_diffueraser/.
• No depth stage in this cache.