aaronyee/hypersim-train-subset

Hypersim train-split subset (cam_00)

A small subset of Apple's Hypersim dataset: 55 scenes from the official train split, camera cam_00 only. Every file is byte-for-byte the file in the official scene archive, at the same path inside the scene.

Source

• Images and labels: the official Hypersim scene archives, https://docs-assets.developer.apple.com/ml-research/datasets/hypersim/v1/scenes/<scene>.zip (release v1), read by HTTP range requests.
• Split: evermotion_dataset/analysis/metadata_images_split_scene_v1.csv in apple/ml-hypersim at commit 3463c5c4a75f3cbfc65ed31cfd6e87204b3a2254.
• Camera list used to draw the first group of scenes: contrib/mikeroberts3000/metadata_camera_parameters.csv at the same commit.

Contents

Folder	Scenes	Files	Content
cam00_full/	15	2,949	For each scene: every cam_00 frame's tonemap.jpg (final preview) and semantic_instance.hdf5; the cam_00 camera files (_detail/cam_00/); _detail/metadata_scene.csv; the semantic-instance bounding boxes (_detail/mesh/metadata_semantic_instance_bounding_box_object_aligned_2d_{extents,orientations,positions}.hdf5)
cam00_depth_semantic/	the same 15	300	10 evenly spaced cam_00 frames per scene: depth_meters.hdf5 and semantic.hdf5
cam00_extra_tonemaps/	40 others	1,600	40 evenly spaced cam_00 frames per scene: tonemap.jpg only
files.tsv			every file: path, bytes, sha256

Below each folder, paths are the paths inside the official archive, for example cam00_full/ai_004_009/images/scene_cam_00_final_preview/frame.0000.tonemap.jpg. A few scenes have fewer frames because the official release leaves some frames out.

Processing

Nothing is converted, re-encoded, resized or relabeled. The only processing is the choice of scenes, frames and files:

1. cam00_full/: 20 scenes were drawn with a fixed seed from the scenes whose camera is the standard pinhole (no lens shift, 60° horizontal field of view, 1024 × 768) in the camera list above. For each, the files listed in the table were read from the scene archive. The 5 scenes that the official split puts in test (ai_005_001, ai_035_001, ai_041_004) or val (ai_046_005, ai_052_001) were then removed, leaving 15.
2. cam00_depth_semantic/: for the same 15 scenes, 10 frames evenly spaced over the scene's available cam_00 frames (seeded offset), and for each the two files listed.
3. cam00_extra_tonemaps/: 40 further train scenes, drawn with a fixed seed from the train scenes that have annotated cam_00 frames in Omni3D's Hypersim training index, excluding scenes already held in a separate full download (every scene of cam00_full/ among them); per scene, 40 frames evenly spaced over those annotated frames. Only the tonemap.jpg files are included. No Omni3D annotation is included.
4. Checked against the official split CSV: all 55 scenes are train, and every per-frame file belongs to a frame marked included_in_public_release = True with split train.

Split

Training data only: all 55 scenes are in Hypersim's official train partition. No val or test scene, and no frame outside the public release, is included.

License

Creative Commons Attribution-ShareAlike 3.0 Unported (CC BY-SA 3.0), the license of the Hypersim Dataset ("The Hypersim Dataset is licensed under the Creative Commons Attribution-ShareAlike 3.0 Unported License", ml-hypersim README).

Share-alike: this subset is distributed under CC BY-SA 3.0. If you remix, transform or build upon it, you must give appropriate credit, link to the license, state what you changed, and distribute your contributions under the same license (CC BY-SA 3.0 or a compatible license).

Changes from the original: selection of scenes, frames and files only, as described above. The files themselves are unchanged.

Attribution and citation

Hypersim was created by Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb and Joshua M. Susskind (Apple). Source: https://github.com/apple/ml-hypersim. If you use this data, please cite:

@inproceedings{roberts:2021,
    author    = {Mike Roberts AND Jason Ramapuram AND Anurag Ranjan AND Atulit Kumar AND
                 Miguel Angel Bautista AND Nathan Paczan AND Russ Webb AND Joshua M. Susskind},
    title     = {{Hypersim}: {A} Photorealistic Synthetic Dataset for Holistic Indoor Scene Understanding},
    booktitle = {International Conference on Computer Vision (ICCV) 2021},
    year      = {2021}
}