ZRY000/test-data-20261023

Dataset structure

Each sample contains one full RGB image, one metric depth image, and one serialized annotation dictionary with four hand entries.

Dataset statistics

The complete dataset contains 855,775 samples: 854,016 samples in 834 TAR shards, plus 1,759 remainder samples stored as individual files. Every TAR contains exactly 1,024 samples and 3,072 member files. One sample is one full image with its depth and annotation dictionary; its four hand entries are part of that sample and do not count as separate samples.

Split	TAR shards	Samples in TARs	Remainder samples	Total samples
train	412	421,888	969	422,857
val	121	123,904	349	124,253
test	301	308,224	441	308,665
Total	834	854,016	1,759	855,775

All TARs, remainder files, and accompanying metadata have been uploaded and verified against their remote content hashes. Samples are shuffled within each split using a large sample pool, and each TAR mixes multiple recordings.

Directory layout

data/
  train/*.tar
  val/*.tar
  test/*.tar
remainders/
  train/
  val/
  test/
metadata/
  coverage_plan.json
  source_attribution.json
  shards/{train,val,test}/*.json
  PROCESSING_COMPLETE.json
preview.jpg

data/ contains complete shards. remainders/ contains the final samples that do not fill a complete shard. metadata/ contains coverage information, per-shard sample ranges, source-frame identifiers, recording distributions, content hashes, and provenance. PROCESSING_COMPLETE.json is added when local coverage checks are complete. preview.jpg is a visualization example.

TAR members

Every TAR contains exactly 1024 samples, with three members per sample:

00000000.data.pyd
00000000.depth.png
00000000.jpg
00000001.data.pyd
00000001.depth.png
00000001.jpg
...

Sample IDs are globally unique eight-digit numbers starting at 00000000. Shard filenames also use eight-digit numbers. Images from multiple recordings are mixed within each shard, while train/val/test directories remain separate.

Component	Format	Contents
.jpg	RGB JPEG, 960×540	One full image
.depth.png	Single-channel uint16 PNG	Depth in millimetres; multiply by 0.001 for metres
.data.pyd	Portable protocol-4 pickle	One schema-v2 annotation dictionary using built-in values

Annotation dictionary

Field	Structure
schema_version	Integer, 2
frame_uid	Original recording/view/frame identifier
image_size	[height, width]
camera	Pinhole intrinsics K and calibration metadata
depth	Unit, metric scale, invalid code, depth type, and provenance
joints_3d_frame	camera
joints_3d_unit	m
annotations_complete	Boolean availability/completeness indicator
hands	List of four hand dictionaries
split	train, val, or test

Each hand dictionary contains:

Field	Shape or type
right	Boolean side indicator
betas	10 shape coefficients
hand_pose	48 axis-angle values: global orientation plus 15 local joint rotations
cam_t	Camera-frame translation, 3 values in metres
keypoints_2d	21×3: pixel coordinates and validity
keypoints_3d	21×4: camera-frame coordinates in metres and validity
has_betas, has_hand_pose	Target availability flags
Additional metadata	Role, subject identifier, parameter provenance, and fitting residuals

Remainder members

Each remainder directory stores the same three components as individual files and a manifest.json. Remainders stay in their original split and are not packed into smaller TARs. Samples are neither dropped nor duplicated or padded to fill a TAR.

Directory	Samples	Sample ID range
remainders/train/	969	00854016–00854984
remainders/val/	349	00854985–00855333
remainders/test/	441	00855334–00855774

Each manifest has split, samples, records, and reason fields. Each entry in records contains a sample_id and its original frame identifier, uid. To load a remainder sample, use its sample_id as the shared filename prefix and read <sample_id>.jpg, <sample_id>.depth.png, and <sample_id>.data.pyd from the same directory. These files use exactly the same formats and annotation schema as the TAR members.

The complete TARs use sample IDs 00000000–00854015; remainder IDs continue through 00855774. To cover all 855,775 samples, read both data/{split}/*.tar and remainders/{split}/manifest.json with its referenced files. Reading only the TARs covers 854,016 samples. Remainder storage contains 5,277 sample files and three split manifests. Final coverage checks confirm that every planned sample appears exactly once across TARs and remainders.