re10k-wds

RealEstate10K (WebDataset)

RealEstate10K camera trajectories and frames, in the pixelSplat chunk release, repacked as random-access WebDataset shards.

Every file in this repository was verified against the source copy (same file set, same byte sizes).

Format

WebDataset tar shards (about 3 GB each), one tar row per scene, indexed for random access with ultar .utix files (msgpack rows of entry keys, offsets and sizes, so they can be read without ultar). Every scene row holds:

Beside each shard:

• <shard>.tar.utix — the ultar row index (offsets and sizes of every entry).
• <shard>.tar.centres.bin — float32 camera centres per frame per scene (a small sidecar used for farthest-point view sampling); header DPFV, then a per-scene offset table in index row order.
• <split>/index.json maps every scene key to its shard; splits/<split>.txt lists the keys; splits/excluded.jsonl lists scenes that could not be converted (empty for this release).

Nothing is re-encoded: frames, cameras and timestamps are byte-identical to the source chunks, so results computed on these shards match results computed on the original .torch chunks.

Source, license and copyright

This repository is a format conversion (WebDataset/ultar) of a publicly released dataset; it adds no new images. Please cite and follow the terms of the original work:

• RealEstate10K: T. Zhou et al., Stereo Magnification: Learning View Synthesis using Multiplane Images, SIGGRAPH 2018 — https://google.github.io/realestate10k/

• Source chunks: the pixelSplat release, mirrored at https://huggingface.co/datasets/lhmd/re10k_torch (pixelSplat: D. Charatan et al., CVPR 2024 — https://github.com/dcharatan/pixelsplat)

• The camera trajectories and video lists are released by the original authors under their own terms (see the links above).

• The frames come from third-party online videos and remain the copyright of their owners. They are provided here, as in the source chunk repository, for non-commercial research use only.
• The conversion code layer (this packaging) is released under the MIT license, matching the source chunk repository.

Removal requests. If you own content in this dataset and want it removed, please open a discussion on this repository; affected scenes will be removed promptly.

Citation

Please cite the original dataset:

@inproceedings{zhou2018stereo,
  title     = {Stereo Magnification: Learning View Synthesis using Multiplane Images},
  author    = {Zhou, Tinghui and Tucker, Richard and Flynn, John and Fyffe, Graham and Snavely, Noah},
  booktitle = {SIGGRAPH},
  year      = {2018}
}