Datasets:

Modalities:
Audio
Video
Size:
> 1T
Libraries:
Datasets
License:
GameRealm / Tracking /Standard /README.md
weiguangzhao's picture
Add GameRealm partial release: Tracking_Standard_1of5
8616219 verified
|
Raw History Blame Contribute Delete
4.26 kB

GameRealm Tracking Standard

This collection is being converted and uploaded. catalog.jsonl contains only fully committed sequences. progress.json reports the current verified totals; an unfinished directory without its _status receipt is not a completed sample.

Only sources whose every camera has original width 1280 and height 720 are admitted. Every decoded RGB and depth frame must independently meet that size. No resize, padding, crop, frame thinning or point sampling is performed. Current benchmark holdout source recordings are excluded. Source recordings and previous releases are preserved outside this publication directory.

Single-view samples are singleview/00001/. Multi-view samples are multiview/00001/cam_00/, cam_01/, etc. Identifiers are opaque and remain fixed; gaps can correspond to pending or failed review. Each view has four files:

  • rgb.mp4: H.264, CRF 18, YUV420, original frame sequence and recorded timing.
  • depth.h5: camera-Z depth in centimetres, log-u16, row delta, byte shuffle, Zstd-3, independent frames and an embedded byte-range index. Float32 exceptions are preserved. _reader/depth_log_codec.py reads this file.
  • meta.h5: frame IDs, timestamps, intrinsics, world-to-camera and c2w matrices, modality validity, and explicit availability flags for controls and audio.
  • trajectory.h5: all original stored points, their world/pixel trajectories, three separate masks, original numeric source fields, and local SE3 where available. Fixed query selection is not imposed: points/query_frame=-1.

points/point_id is the original zero-based point index. Its identity is shared across views of the same sequence. No game names or private source paths are included. Embodied sources retain their dataset identity in source_dataset.

Validity and visibility

The definitions are also embedded in each trajectory file's rules group:

  • masks/valid: source geometry validity, per-view geometry eligibility, finite world coordinates and positive camera Z. Source 3D label/per-view label gates are additionally retained when present. A false value does not necessarily mean that no numerical coordinates exist.
  • masks/visibility: source visibility state equals 1.
  • masks/visibility_known: source visibility state is one of 1, 2, 3 or 4.

Ignore visibility wherever visibility_known is false. Unknown visibility is kept separate from geometric validity. Source visibility is preserved and is not recomputed from newly compressed images or depth. Original source masks and other numeric fields are retained under source_numeric.

Coordinates, timing and missing modalities

Camera matrices and trajectories retain the source coordinate units and values. Depth alone is converted to centimetres. Use camera.input_to_m and depth.input_to_cm for the recorded conversion convention; working engine scale conventions are not independent physical calibration. c2w is the inverse of the stored world-to-camera matrix. Local SE3 retains reference frames, point binding, transforms, support masks and fit diagnostics, without a new ground-truth claim.

Frame timestamps retain their recorded source clock, which can be simulation, capture or nominal delivery time. MP4 timestamps use microsecond ticks; equal timestamps receive a one-microsecond playback increment. Original timestamps and the adjustment count remain in meta.h5. Missing capture timestamps use -1. Controls unavailable in the Standard numeric contract are explicitly absent; no control values are invented. No audio.wav is published; has_audio=false.

Review gates

Before uploading any view of a sequence, all its views must pass source SHA256, numeric schema/mask/camera checks, full video decode and timestamp checks, every-frame depth error/exception checks, exact numeric readback and text-field anonymization. Every uploaded payload then receives a full cloud GET/SHA256 comparison. Only after all files pass is the sequence receipt and catalog entry committed. Conversion checks do not certify independent geometric ground truth.

Native embodied representations that require a dense geometry/SE3 reader remain pending adaptation and are not silently replaced by sampled sparse annotations.