OSResight/hypersim-episodes

Hypersim Episode Pairs

WebDataset shards of frame pairs from Apple Hypersim (CC BY-SA 3.0) for geometry-aware training. Compatible with create_geometry_loader() from kvray-distill.

Dataset Summary

Split	Shards	Pairs	Scenes
Train	72	88,354	208
Val	10	7,255	18
Total	82	95,609	226

• Image size: 1024x768
• Frame gaps: 1, 2, 4, 8 (wider gaps subsampled to 50% of gap=1 count)
• Split provenance: metadata_images_split_scene_v1.csv from apple/ml-hypersim

Shard Schema

Each sample in a shard contains:

File	Description
key.rgb0.jpg	Frame 0 RGB (JPEG quality 85, Reinhard tone-mapped from HDR)
key.rgb1.jpg	Frame 1 RGB
key.depth0.npz	Frame 0 depth (float16, compressed)
key.depth1.npz	Frame 1 depth (float16, compressed)
key.seg0.npz	Frame 0 segmentation (uint16, NYU-40) — when available
key.seg1.npz	Frame 1 segmentation (uint16, NYU-40) — when available
key.meta.json	Poses, intrinsics, gap, metadata

meta.json Fields

{
  "dataset": "hypersim",
  "depth_type": "ray_distance",
  "pose_convention": "T_w_c",
  "scene": "ai_001_001",
  "camera": "cam_00",
  "frame_ids": ["ai_001_001_cam_00_0010", "ai_001_001_cam_00_0011"],
  "gap": 1,
  "poses": [[...16 floats (4x4 row-major)...], [...]],
  "intrinsics": [fx, fy, cx, cy],
  "image_size": [768, 1024],
  "has_flow": false,
  "has_seg": true,
  "meters_per_asset_unit": 0.0254
}

Depth Convention

IMPORTANT: Depth values are ray distance (Euclidean distance from the optical center), NOT projective z-depth.

To convert to projective z-depth for standard pinhole geometry:

dx = (u - cx) / fx
dy = (v - cy) / fy
cos_angle = 1.0 / np.sqrt(1.0 + dx**2 + dy**2)
z_depth = ray_distance * cos_angle

The create_geometry_loader() in kvray-distill handles this automatically when depth_type == "ray_distance" in meta.json.

Pose Convention

Poses are T_w_c (camera-to-world) as 4x4 homogeneous matrices in row-major order. Positions are in meters (converted via per-scene meters_per_asset_unit).

Intrinsics

Derived from M_cam_from_uv in metadata_camera_parameters.csv:

• fx = W / (2 * M_cam_from_uv_00) ≈ 886.8 px
• fy = H / (2 * M_cam_from_uv_11) ≈ 886.8 px
• cx = W / 2 = 512, cy = H / 2 = 384

All Hypersim scenes have settings_camera_fov = pi/3 (60 deg HFOV).

Excluded Scenes

207 tilt-shift scenes excluded where the physical camera model introduces lens shift that a simple pinhole intrinsics model cannot represent (off-diagonal M_cam_from_uv entries > 0.01, corresponding to >5px principal point error). See audit.json for the full exclusion list.

4 pairs with severely broken rotation matrices (orthogonality error > 100) were removed post-build. See audit.json for details.

Quality Audit

Full audit results in audit.json. Summary:

• Tilt-shift contamination: CLEAN
• Depth convention: PASS
• Pose convention: PASS
• Intrinsics range: PASS
• Depth plausibility: PASS
• 78 minor pose orthogonality warnings (floating-point noise, < 0.1% of pairs)

License & Attribution

This dataset is derived from Hypersim by Apple Inc., licensed under CC BY-SA 3.0.

Mike Roberts and Nathan Paczan. "Hypersim: A Photorealistic Synthetic Dataset for Holistic Indoor Scene Understanding." International Conference on Computer Vision (ICCV) 2021.