Datasets:

Modalities:
Audio
Video
Size:
> 1T
Libraries:
Datasets
License:
weiguangzhao's picture
Add GameRealm partial release: Tracking_Benchmark_full
60be9be verified
|
Raw History Blame Contribute Delete
5.86 kB

GameRealm Tracking Benchmark

This release reorganizes the accepted benchmark into 167 single-view sequences and 8 three-view sequences: 175 sequences, 191 camera inputs, 104306 camera frames. The Long-horizon subset remains 20 sequences. Five-digit identifiers are fixed within each subset. The catalog maps them to original benchmark IDs, categories, parent sequences and anonymized synthetic or named embodied dataset sources.

Each singleview/00001/ or multiview/00001/cam_00/ contains exactly four modalities:

  • rgb.mp4: H.264, CRF 18, normally yuv420p, original resolution and source timing. IP-03 through IP-06 now use native 1280x720 sources and yuv420p. See _reports/video_format_notes.json for revision details. RGB encoding is lossy.
  • depth.h5: camera-forward depth, converted to centimetres using the recorded per-source working scale, then log-u16, horizontal delta, byte shuffle and independent Zstd-3 frames. Exceptions preserve float32 values exactly.
  • meta.h5: frames, camera, actions, audio, valid and capture/original groups.
  • trajectory.h5: original world/pixel trajectories, point IDs, query frames, visibility, visibility_known and valid masks, with embedded decision rules.

There is no audio.wav. has_audio=false. Missing actions and capture timestamps are explicitly marked unavailable; no controls or capture timing are invented. Source frame IDs may be strings. timestamp_ns preserves the available source clock; it is not always wall-clock capture time. Some source timestamps are nominal simulator/delivery timestamps; frames.timestamp_clock and timestamp_measured distinguish these. Missing capture timestamps use -1 with a false validity mask. MP4 uses microsecond time ticks; original timestamps are retained. Equal timestamps receive a minimal one-microsecond playback increment.

Camera K and world-to-camera matrices retain frozen input values; c2w is their inverse. This is the benchmark coordinate basis, which may differ from raw capture matrices. Camera translations and world trajectories retain their source units. They must NOT be interpreted as centimetres just because depth is cm. camera.input_to_m and depth input_to_cm record the conversion. Working engine scale conventions are not independent physical scale calibration.

Trajectories and masks

trajectory.h5 contains:

  • frames/frame_id: [T], frame index in this view.
  • points/point_id: [N], preserved source point identifier.
  • points/query_frame: [N], original zero-based query frame.
  • tracks/xyz_world: [T,N,3], source world coordinates, losslessly preserved.
  • tracks/uv: [T,N,2], source pixel coordinates, losslessly preserved.
  • masks/valid: [T,N], original geometric/evaluation eligibility.
  • masks/visibility: [T,N], visible only where visibility_known is true.
  • masks/visibility_known: [T,N], whether visibility is classified.
  • rules/: attributes defining the source policy, criteria and thresholds.
  • provenance/source_state: preserved source state codes for auditability.

Ignore visibility when visibility_known=false. Do not merge unknown visibility into geometric validity. valid is the original source's eligibility mask: it may include positive camera depth and source label gates, so false does not prove that a numerical point position is missing. A consumer needing both valid geometry and classified visibility may use valid & visibility_known.

Visibility labels are preserved from the sealed benchmark and are not recomputed against newly quantized depth or re-encoded RGB. Published canonical labels and depth-interval-derived labels retain their respective origins. For the latter, the conservative point/depth ratio interval must fit inside [0.99,1.01] to be visible; a lower bound above 1.01 is occluded. Unsupported/ambiguous depth and uncertain intervals remain unknown. See each trajectory file's source rules.

Multiview sequences have matching source frame IDs and source timestamps across their included cameras. Source clocks are not shared across different sequences. Cross-camera point correspondences are not asserted: each view retains its own frozen query selection. Equal point positions in an array do not imply identity.

Reading depth

Use _reader/depth_log_codec.py (Python, NumPy, h5py, zstandard):

from depth_log_codec import DepthLogReader
depth_cm = DepthLogReader('singleview/00001/depth.h5').read(0)

Depth's byte-range index is embedded in depth_log_index_json; no sidecar is needed. Camera matrices and trajectory datasets can be read directly with h5py.

Verification and completion

Each view is uploaded only after full video decode/frame/PTS verification, every depth frame's quantization and exception checks, and exact label/camera readback comparisons. Every uploaded file then passes full cloud GET/SHA256. _status/views records individual views; _status/singleview and _status/multiview commit entire sequences. complete.json is emitted only after all 191 active inputs pass. An incomplete directory without its final receipt is not a finished item.

This format uses lossy RGB and depth quantization and is numerically different from the old NPZ release. Old benchmark inputs and the previous cloud release are retained outside this destination for reproducibility. No raw mesh, private source paths or credentials are included. Hugging Face publication, licensing and public access settings remain separate pending decisions.

Plant revision, 2026-11-15

IP-03 through IP-06 were replaced with four approved native 1280x720 clips, each using 241 consecutive frames. Plant remains 15 sequences / 3690 frames. The benchmark now has 42173 query points. All affected model scores are being recomputed; historical scores for these four inputs do not apply to this revision. Previous source recordings and payloads are retained in the prior release history.