wangchenlong/training_data_public_v2

SpatialEncoder WDS shards

Dynamic Replica and Virtual KITTI 2 in spatialencoder-wds-native-v1: uncompressed WebDataset tar shards of about 1 GiB, each with a JSON sidecar. Sample layout is <key>.meta.json followed by <key>.assetNNNNN.<ext>. meta.json holds the native annotation and maps source paths to assets. Video blocks own 64 clip starts plus up to 14 lookahead frames, so frames in manifest.json counts overlap. urls/<split>.txt lists shard URLs pinned to their upload commit. fingerprints in each sidecar are SHA-256 over the meta JSON and asset bytes, in tar order.

Dataset	Splits	Depth	default_scale	3D boxes
DynamicReplica	train, val (source valid), test	uint16 PNG in mm, 0 = invalid	1000	none, objects is empty
VKITTI2	train (no official split)	uint16 PNG in cm, 65535 = sky	100	native

Changes from the source releases

• Dynamic Replica RGB is re-encoded from RGBA PNG to JPEG quality 95, and depth from float16 metres to uint16 millimetres. Its real split, flow, trajectories, masks and instance maps are not included.
• VKITTI2 RGB is JPEG. Boxes were converted to camera-frame centre, extent and rotation.

Sources and licenses

Source datasets keep their own licenses and access terms. This format grants no additional rights.

• Dynamic Replica: Karaev et al., "DynamicStereo: Consistent Dynamic Depth from Stereo Videos", CVPR
  2023. https://dynamic-stereo.github.io/
• Virtual KITTI 2: Cabon et al., "Virtual KITTI 2", 2020. CC BY-NC-SA 3.0. https://europe.naverlabs.com/research/proxy-virtual-worlds/