RETRACK3D

Test trial for dataset export and dataloader validation. Not an official release.

Storage and continuity

loader/                         standalone Python reader
examples/read_pair.py           read 96 + 96 as a 192-frame sample
docs/FORMAT.md                  schema, identities, coordinates and sampling
validation/                     parity, data checks and training reports
integration/                    InfD4RT adapter and resumable training entry point
xperience/chunks96/v2/           test data payload
  manifest.json
  manifest.sha256
  chunks/<capture>_ep<n>/
    0000000-0000096-<hash>.h5
    0000096-0000192-<hash>.h5
    ...

Frame intervals are half-open. Adjacent files retain episode-global frame IDs, object-track IDs and the same original world coordinate system. The reader concatenates raw arrays before sampling queries and estimating gravity once for the full window. Tail files contain 96–191 frames to avoid padding or dropping a short tail. Whole episodes shorter than 96 frames are preserved too; configure a shorter sample window to use episodes shorter than the default 48 frames. Images use H.264 CRF 18; depth, geometry, masks and IMU use lossless compression at their source/export precision.

Each supervision head has an independent validity mask. Unknown visibility is not an invisible label. Static background targets through occlusion follow a rigidity assumption; these are pseudo-labels, not independent hidden-point GT.

Reader

Install loader/requirements.txt; add loader/ to PYTHONPATH, or run examples/read_pair.py, which finds the bundled loader automatically. The reader accepts local paths, HTTP and fsspec URIs, including Hugging Face. It fetches only the manifest at initialization and shares a checksum-verified, bounded disk cache between workers on each node. Warm cache hits download zero bytes.

With an authorized local payload:

python examples/read_pair.py --manifest /path/to/manifest.json \
  --episode 0073a4be-ec41-47f3-92d6-f63cf89aec9b_ep8.h5 --start 960

See FORMAT.md for all fields and the InfD4RT training interface. Training checks exercised the 1.535B-parameter model at 48 frames / 224 pixels on an A40, including checkpoint resume. A separate small model trained on joined 192-frame inputs. All tested steps had finite losses/gradients and actual parameter updates; they do not establish convergence or annotation accuracy.

Source and terms

Source: Xperience-10M, Ropedia. Its data card specifies controlled access and an external agreement. Underlying source-data rights and terms continue to apply; this repository does not relicense those recordings under MIT. The payload is distributed under manual gated access.

Test export

The current test export is under xperience/chunks96/v2/. The v2-preview directory contains the earlier test subset. Windows and queries are sampled at runtime, including across adjacent files.

Remote test reader

Use hf://datasets/2inf/RETRACK3D/xperience/chunks96/v2/manifest.json with an approved Hugging Face account. Pin the repository revision for reproducible training. The default query allocation is background 50%, object-track 35%, hand 15%; missing pools redistribute the quota without weakening validity masks. Object-track type is not a guarantee of independent motion or a manipulated object. See validation/full_upload.json for upload integrity checks.