Panoptic Pairwise Distortion Graph ✨ [ICLR 2026] ✨

---

Overview

The core idea is to represent comparative image quality as a structured graph over paired images:

• Objects / regions correspond to matched semantic regions in an anchor image and a target image.
• Attributes capture region-level distortion type, severity, and score.
• ART relations capture whether a region in the anchor is the same / slightly better / significantly better / slightly worse / significantly worse than its counterpart in the target.
• Scene relationships can optionally be attached from panoptic scene graph annotations when available.

---

Repository Structure

.
├── config.yml                # Main experiment configuration
├── train.py                  # Distributed training + validation/testing
├── pandadg.py                # Model definition
├── loaddata.py               # PandaBench dataset loader + collate fns
├── helper.py                 # Label maps, metrics helpers, utilities
├── non_graph_inference.py    # Standard evaluation script (acc/F1/SRCC/PLCC)
├── graph_inference.py        # Export predictions as distortion-graph JSON
├── plot_graph.py             # Render saved graph JSON with Graphviz
└── pandabench_idx.py         # Cached easy / medium test split indices

---

Data

Download the dataset (~27GB) from: Hugging Face Link

The default configuration expects the dataset under:

data/pandabench/

with degradation-specific folders plus metadata/statistics JSON files. The config also expects per-degradation statistics files such as:

• train_stats_mixed.json
• train_stats_snow.json
• train_stats_brightness.json
• ...

and swaps train -> val / test automatically inside the loader depending on split.

The expected directory pattern is effectively:

data/pandabench/
├── gt/
│   ├── train/
│   ├── val/
│   └── test/
├── mixed/
│   ├── train/
│   ├── val/
│   └── test/
├── mixed2/
├── snow/
├── brightness/
├── contrast_inc/
├── contrast_dec/
├── compression/
├── saturate_inc/
├── saturate_dec/
├── oversharpen/
├── blur/
├── noise/
├── haze/
├── rain/
├── pixelate/
├── darken/
├── depth/
└── stats/

---

Training

Training is launched through train.py, which initializes distributed training, builds a DeepSpeed config on the fly, creates an experiment directory under ckpts/<experiment_name>/, and logs TensorBoard runs to runs/<experiment_name>/.

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

During training / validation / test, the code tracks metrics such as:

• comparison accuracy and recall@2,
• anchor/target distortion accuracy,
• anchor/target severity accuracy and recall@2,
• anchor/target score MAE,
• validation loss terms.

Note: These metrics are just signals of training progressing and are not the metrics reported/used.

Checkpoints

Checkpoints are saved to: ckpts/<experiment_name>/, with periodic saves like: ep_10.pth (control the saving epochs in config) and a final checkpoint like: final_ep_<N>.pth. Each checkpoint stores epoch, model state, optimizer state, and loss.

---

Inference

Use non_graph_inference.py to run evaluation and print numeric metrics. These are the metrics reported/used.

python non_graph_inference.py --configpath config.yml

This script loads the configured checkpoint, runs inference on the test split. Define the split of the testset in the config.yml file.

---

Graph Export

Use graph_inference.py to convert model predictions into structured distortion graph JSON files.

python graph_inference.py --configpath config.yml

For each evaluated image pair, the script writes a JSON file to:

inf_graphs/

with filenames of the form:

<idx>_<img_tag>_<anchor_deg>_<target_deg>_<experiment_name>.json

Graph Schema

The exported JSON uses four top-level fields:

• objects
• attributes
• relationships
• art

Conceptually:

• objects represent corresponding regions in anchor and target images,
• attributes store distortion / severity / score predictions for each region,
• relationships store within-image scene graph predicates when available,
• art stores cross-image comparative relations between matched regions.

---

Visualization

After exporting graph JSON files, you can render them with Graphviz using plot_graph.py.

python plot_graph.py

The plotting script reads a graph JSON file from inf_graphs/, builds a colored node-edge visualization, and saves a PNG using Graphviz. The default script uses a hard-coded input path, so you will usually want to edit the path variable first.

---

Acknowledgement

We thank the authors of following works for open-sourcing their work which allowed us to build PandaSet/PandaBench.

• DepictQA
• OneRestore
• Open-PSG
• Seagull

---

Citation

If you use this codebase in your research, please cite the Distortion Graph paper.

@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}
}