Infinigen Stereo Video Dataset (InfinigenSV)

Dataset Summary

InfinigenSV is a synthetic stereo video dataset constructed using the photorealistic 3D environment engine Infinigen. It contains high-resolution stereo video sequences across diverse natural outdoor scenes such as forests, mountains, rivers, and rocky terrains, simulating real-world complexity with controlled camera motion and lighting.

The dataset is designed to support research on stereo video matching, temporal consistency, and 3D scene understanding, especially under challenging conditions like textureless surfaces, fine geometry, and dynamic parallax.

Ground-truth depth maps, along with full camera intrinsics/extrinsics, are provided per frame.

Dataset Structure

The folder structure of the dataset is organized by scene categories and sequences:

/data/datasets/InfinigenSV/
  ├── train/
  │   ├── 2873f0f/
      │   ├── frames/
      │   │   ├── Image/
              │   ├── camera_0/
          │   │   │   ├── Image_0_0_0001_0.npy
          │   │   │   ├── Image_0_0_0002_0.npy
              │   ├── camera_1/
      │   │   ├── Depth/
              │   ├── camera_0/
          │   │   │   ├── Depth_0_0_0001_0.npy
          │   │   │   ├── Depth_0_0_0002_0.npy
              │   ├── camera_1/
      │   │   └── camview/
              │   ├── camera_0/
          │   │   │   ├── camview_0_0_0001_0.npz
          │   │   │   ├── camview_0_0_0002_0.npz
              │   ├── camera_1/
  ├── val/
  ├── test/
  └── ...

Data Instances

Each data instance corresponds to a single stereo frame in a video sequence and includes the following fields:

Field	Type	Description
left	RGB image	Left view of the stereo pair (720×1280×3)
right	RGB image	Right view of the stereo pair (720×1280×3)
depth	depth file	Depth in meters (720×1280×3)
camview	camera file	Camera intrinsics and extrinsics per frame

License

The code is under MIT license. Infinigen SV dataset is licensed under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0).

Citation Information

@article{jing2024match,
  title={Match Stereo Videos via Bidirectional Alignment},
  author={Junpeng Jing and Ye Mao and Anlan Qiu and Krystian Mikolajczyk},
  year={2024}
}