Datasets:

---

ParityLab

/

Hand-Datasets-Raw

Like 0

Follow

ParityLab 9

Candidate PR: add EPIC-KITCHENS-VISOR/shared-v1.1, the verified
hand-warehouse v1.1 release of EPIC-KITCHENS VISOR. Please review before merging.

This is the only dataset in the batch whose v1.1 release is not a migration. VISOR never
entered this repository as a v1.0 -- main carries only EPIC-KITCHENS-VISOR/raw -- so
there is nothing to rename, nothing to copy server-side and nothing to delete. Every file
below is new, and the only path outside the new directory this PR writes is one
.gitattributes line.

What this PR adds

35,726 files, 20,364,268,432 B, all under EPIC-KITCHENS-VISOR/shared-v1.1/:

• data/ -- 35,668 files, 179 views (115 train / 43 val / 21 test), 50,729 frames, one
  ego camera per sequence, 34,162 object directories, 367,564 polygon rings and
  21,116,613 points. Every annotated frame carries polygons in source pixel coordinates
  plus contact relations; the 21 test views publish observation only, because the source
  withholds their annotation.
• script/visor_v1_1/ -- 18 files: the release's own converter, verifier, reader,
  validator, renderer and the spec bundle it was built against.
• doc/ -- 40 files: the v1.1 contract, the migration record, the raw-to-shared and
  source-semantics notes, the schema, the full migration summary, the v1.0 semantics
  contract, and a 14-page manual review set under doc/validation/.

Media is the source JPEG codestream re-wrapped into .mkv without re-encoding: 179 files,
19,220,018,281 B. The release has no numeric tensor bundle at all -- the only .pt
files are 495 per-frame boolean validity files (rgb_valid.pt, segment/mask_valid.pt,
contact/relations_valid.pt), 912,535 B in total. There is no per-frame numeric component
for v1.1's merge rules to bind.

Why there is no migration step here

MIGRATION.md assumes a dataset already has a shared-v1.0 to upgrade in place. VISOR does
not, and the five steps that assume one -- lock, stage, atomic replacement of old files,
removal of the old tree, and pinning the source to a previous release's commit -- have no
object here. The deviation is registered in doc/PUBLICATION_PLAN.md rather than left
implicit, and the source is pinned a different way: convert.py asserts the identity of the
published raw archive (30,572,126,007 B, sha256
9e15bb03b408acd28b5ef54f1b752c18909fd995464f1267dc3081711c2e0a7c) before it reads
anything, and raw/doc/archive_checksums.json in the already-published raw tree records the
same digest. EPIC-KITCHENS-VISOR/raw is untouched by this PR.

doc/V1_0_SEMANTICS.md

Shipped, and deliberately: the v1.1 contract's ninth line defers every unchanged semantic to
"the v1.0 contract", and that document is the warehouse-level convention -- not this
dataset's material. It is 45,148 B, sha256
c34cdf6645f91233cc28d86f8abc82f019f78767d863e225f75722984d9adb19, byte-identical to the
copy ARCTIC, RHD, EgoHOS and ReInterHand each publish. Because this dataset has no v1.0 tree
to copy it from, the release's doc builder checks it against that digest at build time and
exits if it differs. What this dataset genuinely does not have is the two things that
preserve an old tree -- doc/history-v1.0/ and script/v1_0_snapshot/ -- and the summary
records exactly that, rather than claiming nothing was carried over at all.

.gitattributes

One line added: EPIC-KITCHENS-VISOR/shared-v1.1/**/*.mkv filter=lfs diff=lfs merge=lfs -text.
This repository has no global *.mkv pattern (unlike *.pt, *.zip, *.mp4), so each
dataset with media carries its own glob -- ARCTIC, RHD, EgoHOS and ReInterHand each do, and
so does dex-ycb/shared-v1.1/**/*.obj. At main the file is 6,535,997 B against the Hub's
10 MiB ceiling, so there are about 3.9 MB of headroom: this is not a size rescue, it is the
shape the repository already uses. The glob goes in the PR's first commit on purpose, so the
Hub has no per-path line to append for the 179 .mkv files. Every non-VISOR line is
byte-identical, re-checked on the candidate branch before merging.

Registered decisions the reviewer should look at

These are recorded in script/visor_v1_1/README.md, doc/RAW_TO_SHARED.md and
doc/SOURCE_SEMANTICS.md, and each is a single-place change if you would rather have it
differently:

1. Dense interpolations are excluded (Interpolations-DenseAnnotations/, 158 archives,
   9.76 GiB): a different semantic object (no exhaustive, no contact, no hand side,
   interpolation segment ids, and 854x480 annotations against the sparse set's 1920x1080,
   with no frames in this deposit).
2. Geometry is polygons, not rasterised mask.mkv -- the contract allows either, and the
   source is float polygons; rasterising would introduce approximation the source does not
   have.
3. Frames are ordered by source frame number (108 of 158 source segments are not
   ascending); the source array index is kept per frame in source_annotation_index.
4. Gloves are registered as objects, not mapped onto the fixed hand slots.
5. The licence is recorded, not resolved: the source states three different things
   (CC BY-NC 4.0, the Non-Commercial Government Licence, and a research licence agreement in
   README.txt), and all three are recorded as they are.
6. glove-not-in-contact maps to negative -- the fourth keyword the source uses but
   its own documentation does not list; the mapping table is the single place to change.
7. Entity ids carry a name digest, because three entity-name pairs collapse to the same
   slug within one segment.
8. This is a full release, not a hand-only subset: all 1,280 entity names go up.
9. No camera calibration exists in the source, so camera.model is unknown and the
   review pages overlay pixel-space polygons rather than projecting 3D geometry.

Acceptance, and what is still pending

The release was built and accepted by the eight-step order its README prescribes (convert,
verify against the raw archive, read sweep, structural validation, review-page render,
summary, doc rebuild, final validation with manifest rebuild). The final validation passes
with coverage: full release, unreferenced_files: [], and a rewrite check on two views;
doc/FULL_MIGRATION_SUMMARY.json and doc/validation/validate_report.json carry the
measured numbers.

The human review has not happened yet, and this PR does not claim otherwise. Every
view's metainfo.json carries upload_authorized: false, the summary records
human_full_review: pending and publication_authorized: false, and no write to this
repository is authorized by the release itself. doc/validation/index.html is the material
for that review: 14 sampled views (seed 20261017, stratified by split, plus coverage
additions), each page showing decoded frames with the source polygons and contact lines.

Verification performed on the candidate branch before this was pushed

Exact file set against the release's own manifest; per-file size and content identity (LFS
sha256 for LFS paths, git blob id for plain git paths); EPIC-KITCHENS-VISOR/raw tree id
unchanged; no other dataset's root changed; no non-VISOR LFS line changed; .gitattributes
under the Hub's ceiling. The Hub's 1,000,000-file limit is left for the Hub to enforce: this
repository's file count cannot be walked in a useful time, so a crossing would surface as the
Hub refusing a commit and naming the count, which this tool detects by name and stops on
before merge.

Pushed under yhl's authorization. Merging is left to the reviewer -- the
tool that pushed this branch has no merge option at all, deliberately.

🤖 Generated with Claude Code

User-authorized merge after output-isolation repair. The converter report and every renderer target are preflighted; JSON/JSONL/HTML use unique-temp atomic replacement. Twelve shipped isolation tests and 26 integration scenarios pass, and the committed tree matches the audited candidate exactly. Data/raw/schema and other datasets are unchanged. Historical full-data evidence remains historical, and human review remains pending; this merge is not visual acceptance or a resolution of source-license questions.