Panoptic Pairwise Distortion Graph

Muhammad Kamran Janjua, Abdul Wahab, Bahador Rashidi

✨ International Conference on Learning Representations (ICLR), 2026 ✨

📄 Paper 💻 Code 📃 Poster 🤗 Dataset 🤗 Models

Overview

Problem

Most comparative assessment methods operate at the whole-image level, while implicitly requiring region-level understanding. This becomes a bottleneck for fine-grained reasoning about where (and how) degradations differ across two images.

Core idea

Represent an image pair as a structured composition of matched regions, with edges capturing comparative relations and nodes capturing distortion attributes (type, severity, quality score). This is the Distortion Graph (DG) formalism.

**DG Task Overview:** Top: Given two images, PANDA learns the proposed Distortion Graph (DG). Bottom: Grounded subgraphs illustrate how DG grounds regions in terms of distortion information.

Distortion Graph (DG)

What DG represents

DG is a region-grounded topology over an image pair: nodes correspond to regions (with masks), and inter-image edges encode how the anchor region compares to its matched target region.

What DG stores per region

Distortion Type (e.g., Noise, Blur, Rain)
Severity Level (e.g., Minor, Moderate)
Quality Score (e.g., 0.10, 0.34, 0.78)
Comparative Relation (e.g., Slightly Better/Worse)

Motivation figure — **Motivation:** Current MLLMs struggle with region-level understanding even with explicit region details. DG grounds assessment in regions with structured distortion attributes and relations.

Dataset & Benchmark

PandaSet

A region-level dataset built to supervise DG prediction, spanning distortion types, severity, and region quality.

PandaBench

Benchmark splits (Easy / Medium / Hard) derived from PandaSet to evaluate region-wise comparative reasoning.

PandaSet distortions — **PandaSet Distortions:** Examples of distortion types applied to the same image; some degrade perceptual quality more than others.

PandaBench samples — **PandaBench Samples:** Easy → single distortion; Medium → mixed region-wise; Hard → region-wise variations across both images.

BibTeX

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