PandaBench

Paper | Project Page | Code

PandaBench (and PandaSet) is an image distortion benchmark designed for evaluating perceptual comparison and distortion-aware visual reasoning. It introduces the task of learning a Distortion Graph (DG), representing dense degradation information such as distortion type, severity, and quality scores in a compact, interpretable graph structure grounded in image regions.

PandaSet (the train/val folders) is used to train the model (Panda), while different selections from the test/ folder form the PandaBench. Please refer to the codebase and paper for details.

Dataset Structure

Image folders

Each top-level directory corresponds to a distortion family:

• blur, brightness, compression, contrast_dec, contrast_inc, darken, depth, gt, haze, mixed, mixed2, noise, oversharpen, pixelate, rain, saturate_dec, saturate_inc, snow

Inside each distortion folder, images are split into:

• train
• val
• test

Metadata

The stats/ directory contains JSON files used by the codebase as the source of truth for metadata. Examples include:

• train_stats_blur.json
• val_stats_blur.json
• test_stats_blur.json

Note: hf_preview is not used in the code and should be deleted once the dataset is downloaded. It is provided for aesthetic purposes for the Hugging Face dataset card.

Usage

For detailed instructions, please visit the official GitHub repository.

Training

Training can be launched through train.py:

torch --standalone --nproc_per_node=8 train.py --name "pandadg_run01" --configpath config.yml

Inference

Use non_graph_inference.py to run evaluation and print numeric metrics:

python non_graph_inference.py --configpath config.yml

To export predictions as structured distortion graph JSON files:

python graph_inference.py --configpath config.yml

Citation

If you use this dataset in your research, please cite:

@inproceedings{janjua2026panoptic,
  title={Panoptic Pairwise Distortion Graph},
  author={Muhammad Kamran Janjua and Abdul Wahab and Bahador Rashidi},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2026}
}