infinigen/2023-10-nature-preview

Infinigen 2023-10 Nature Preview

Pre-generated video data from Infinigen, a procedural generator of photorealistic 3D natural worlds. This is the 2023_10_13_preview release, previously distributed from infinigen-data.cs.princeton.edu; the files here are byte-identical to that release.

Preliminary test video release. This release contains only a small sample of 10 videos. Please stay tuned for more, and post any feedback on our GitHub Issues page. InstanceSegmentation and TagSegmentation masks are not included in this release; they will be included in future releases.

Scenes	10
Cameras	2 (camera_0, camera_1; a stereo pair)
Frames per camera	192 (numbered 0001–0192)
Images	3,840
Ground-truth passes	27
Resolution	1280x720 (Depth_npy / Depth_png are 2560x1440)
Total size	249 GB, ~25 GB per scene
Infinigen version	1.2_rc_0
Data-release format version	0.2.0

Downloading

Download only the passes you need. All 27 passes for all 10 scenes is 249 GB; RGB plus depth, normals and camera poses for one scene is under 8 GB.

from huggingface_hub import snapshot_download

snapshot_download(
    "infinigen/2023-10-nature-preview",
    repo_type="dataset",
    local_dir="my_download",
    allow_patterns=[
        "metadata.json",
        "renders/1a1f2004/*_Image_png_camera_0.tar.gz",
        "renders/1a1f2004/*_Depth_npy_camera_0.tar.gz",
        "renders/1a1f2004/*_camview_npz_camera_0.tar.gz",
    ],
)

Or from the command line:

hf download infinigen/2023-10-nature-preview --repo-type dataset --local-dir my_download \
    --include "renders/1a1f2004/*_Image_png_camera_0.tar.gz"

Then extract the tarballs in place:

cd my_download
find renders -name '*.tar.gz' -exec tar -xzf {} -C . \;

Repository layout

metadata.json                                      # scene / camera / pass manifest
renders/{scene}/{scene}_{Pass}_camera_{c}.tar.gz   # 540 tarballs = 10 scenes x 27 passes x 2 cameras
thumbnails/{scene}_thumbnail.png                   # one preview image per scene

Each tarball holds all 192 frames of one pass for one camera of one scene, and extracts to:

{scene}/frames/{Pass}/camera_{c}/{Pass}_0_0_{frame:04d}_{c}.{ext}

The two numeric fields after the pass name are the camera-rig index and the resample index, both 0 throughout this release; the trailing field is the sub-camera, matching the camera_{c} directory.

Scene seeds: 1a1f2004, 1d8d6f8e, 1f25974d, 2d2c1104, 2e962781, 4bbdd3e0, 5be1f943, 22ca3b8b, 560a8258, f020c43.

Ground-truth passes

Sizes are totals across all 10 scenes and both cameras.

Pass	Size	Description
Image_png	6.2 GB	RGB image as a .png.
Image_exr	40.8 GB	RGB image as an .exr. Full dynamic range, before tonemapping.
camview_npz	0.02 GB	Camera intrinsic & extrinsic matrices, i.e. camera calibration and poses.
Depth_npy	40.2 GB	Depth.
Depth_png	3.9 GB	Color-mapped PNG of Depth_npy. FOR VISUALIZATION ONLY.
SurfaceNormal_npy	35.3 GB	Surface normals.
SurfaceNormal_png	7.0 GB	Color-mapped PNG of SurfaceNormal_npy. FOR VISUALIZATION ONLY.
Flow3D_npy	29.5 GB	Optical flow and depth change.
Flow3D_png	1.0 GB	Color-wheel visualization of the 2D part of Flow3D_npy. FOR VISUALIZATION ONLY.
Flow3DMask_png	0.4 GB	Flow occlusion mask.
OcclusionBoundaries_png	3.1 GB	Occlusion boundaries.
ObjectSegmentation_npz	0.2 GB	Semantic segmentation mask, compressed using a lookup table.
ObjectSegmentation_png	0.4 GB	Color-mapped PNG of ObjectSegmentation_npz. FOR VISUALIZATION ONLY.
Objects_json	36.3 GB	Names, poses and bounding boxes of objects in the scene. Required for 2D/3D bounding boxes.
AO_png	3.6 GB	Ambient occlusion.
DiffCol_png	5.0 GB	Diffuse color, a.k.a. albedo.
DiffDir_png	5.6 GB	Diffuse direct pass.
DiffInd_png	5.7 GB	Diffuse indirect pass.
Emit_png	1.3 GB	Emission pass.
Env_png	1.5 GB	Environment pass.
GlossCol_png	3.0 GB	Glossy color.
GlossDir_png	4.8 GB	Glossy direct pass.
GlossInd_png	5.5 GB	Glossy indirect pass.
TransCol_png	1.4 GB	Transmission color.
TransDir_png	1.3 GB	Transmission direct pass.
TransInd_png	2.0 GB	Transmission indirect pass.
VolumeDir_png	3.9 GB	Volume direct pass.

Where both a .png and a .npy/.npz file exist for the same quantity, use the .png only for visualization and the array file for training.

Data formats

Full documentation is in GroundTruthAnnotations.md. In brief, with H and W the height and width of the RGB image:

• Depth — 2H x 2W float32 array.
• Surface normals — H x W x 3 float32 array, coordinate system +X right, +Y up, +Z backward.
• Occlusion boundaries — >= 2H x 2W png, 255 on a boundary and 0 otherwise.
• Flow3D — H x W x 3 float32 array. Channels 1 and 2 are standard optical flow, in pixels at the resolution of the original image; channel 3 is the depth change to the next frame. Values are not meaningful on the final frame of a sequence.
• Flow3DMask — H x W png, 255 for a co-visible pixel and 0 for an occluded one. Computed by comparing triangle face-ids at either end of each flow vector, which yields frequent false negatives distributed like salt-and-pepper noise; max-pooling the mask down to image resolution reduces them.
• camview — .npz with K, the 3x3 intrinsic matrix (pinhole model); T, the 4x4 camera-to-world extrinsic matrix; and HW, the image height and width. The world coordinate system is the computer-vision convention, +X right, +Y down, +Z forward — not Blender's internal convention, which flips Y and Z.
• ObjectSegmentation — H x W int32 array, index-compressed into the .npz as vals (the distinct object indices), indices (one uint8 lookup per pixel) and shape. Reconstruct with vals[indices].reshape(shape). Each integer maps to the entry in Objects_json with the same object_index. "Object" is whatever Blender treats as one object: large or complex assets such as terrain, trees or animals are a single object, while many small instanced assets (grass, coral) may be grouped together.
• Objects_json — per object: object_index, name, children, num_verts, num_faces, materials, unapplied_modifiers, tags, bounding-box corners min and max in object coordinates, and model_matrices, the 4x4 object-to-world transforms of every instance. A 3D bounding box per instance follows from min, max and model_matrices.

Loading

Infinigen ships a PyTorch-style dataset class, dataset_loader.py:

from infinigen.tools.dataset_loader import get_infinigen_dataset

dataset = get_infinigen_dataset("my_download", data_types=["Image_png", "Depth_npy"])
print(len(dataset))
print(dataset[0].keys())

dataset_loader.py is designed to be separable from the main Infinigen codebase — you can copy it into your own project, but you must also copy its dependency suffixes.py.

License

BSD 3-Clause. Copyright (c) 2023, Princeton University.

Citation

@inproceedings{infinigen2023infinite,
  title={Infinite Photorealistic Worlds Using Procedural Generation},
  author={Raistrick, Alexander and Lipson, Lahav and Ma, Zeyu and Mei, Lingjie and Wang, Mingzhe and Zuo, Yiming and Kayan, Karhan and Wen, Hongyu and Han, Beining and Wang, Yihan and Newell, Alejandro and Law, Hei and Goyal, Ankit and Yang, Kaiyu and Deng, Jia},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={12630--12641},
  year={2023}
}