hmorimitsu/resolution-calibrated-corruptions

RCC Dataset

Resolution-Calibrated Corruptions (RCC) is a synthetic image dataset designed for evaluating the effect of corruptions at multiple scales. It was introduced in the Electronics 2026 paper High-Resolution Perception at Scale: Benchmarking Dense Geometric Models via Resolution-Calibrated Corruptions to benchmark optical flow and depth estimation models. It extends the Kubric-NK dataset by providing 22,680 images distributed across 630 samples at 1K, 2K, and 4K resolutions with 12 types of corruptions:

• Blur: Defocus Blur, Gaussian Blur, and Motion Blur.
• Lighting: Low Light, Overexposure, and Sunlight.
• Noise: Gaussian Noise, Shot Noise, and Speckle Noise.
• Weather: Fog, Rain, and Snow.

🚀 How to Use the Dataset

The dataset contains 3 .tar files, containing all the samples for 1K, 2K, and 4K resolutions. Each .tar file contains the images for 30 scenes with 21 frames each exposed to 12 corruptions (7,560 images per resolution).

Important: This RCC dataset provides the exact same images as Kubric-NK, but after applying image corruptions. It does not include ground-truth labels. Therefore, it must be paired with the original Kubric-NK dataset to obtain the respective clean images, optical flow maps, depth maps, and configuration files for specific tasks.

1. Download and Extract

Download the archives for the resolution you want to test (e.g., 1k.tar). Extract them into a single root directory.

• You can check the HF Docs about Downloading Datasets.

2. Expected Directory Structure

After extraction, your folder structure will look like this:

1k/
├── defocus_blur/
│   ├── 000/
│   │   ├── rgba_00000.png
│   │   ├── rgba_00001.png
│   │   └── ...
│   ├── 001/
│   └── ...
├── fog/
│   ├── 000/
│   └── ...

📜 Licensing & Acknowledgements

To satisfy both the software and asset requirements of the generation pipeline, this repository operates under the following licenses:

• Dataset (Images): The corrupted image sequences in this repository are released under the CC BY 4.0 license.
• Generation Code: The base 3D scenes were generated using a customized fork of Kubric (originally developed by Google Research).
• 2D Image Corruptions: Some 2D image corruptions were generated by adapting imagecorruptions to multi-scale, as explained in the paper.

🎓 Citation

If you use the RCC dataset in your research, please consider citing our Electronics paper:

@article{Morimitsu2026HighResolutionPerception,
  title = {High-Resolution Perception at Scale: Benchmarking Dense Geometric Models via Resolution-Calibrated Corruptions},
  author = {Morimitsu, Henrique and Yang, Chun},
  journal = {Electronics},
  volume = {15},
  number = {19},
  year = {2026},
  doi = {10.3390/electronics15194449},
}