MatRes: Zero-Shot Test-Time Model Adaptation for Simultaneous Matching and Restoration

2026-04-11

AI Summary

This work addresses the challenge that in real-world scenarios, image pairs often suffer from severe degradation and large viewpoint variations simultaneously, causing mutual interference between image restoration and geometric matching tasks. To tackle this issue, the authors propose MatRes, a framework that achieves, for the first time, zero-shot joint optimization at test time without offline training or additional supervision. Given only a single pair of low-quality and high-quality images, MatRes leverages positional correspondence through a conditional similarity constraint to jointly refine restoration quality and correspondence estimation. While keeping the pre-trained backbone frozen, it updates only lightweight, adaptable modules. Experiments demonstrate that this approach significantly outperforms existing methods that handle restoration and matching independently across diverse image pair configurations.

Technology Category

• Computer Vision: Multi-modal Vision
• Machine Learning: Multimodal Learning
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs

Abstract

Real-world image pairs often exhibit both severe degradations and large viewpoint changes, making image restoration and geometric matching mutually interfering tasks when treated independently. In this work, we propose MatRes, a zero-shot test-time adaptation framework that jointly improves restoration quality and correspondence estimation using only a single low-quality and high-quality image pair. By enforcing conditional similarity at corresponding locations, MatRes updates only lightweight modules while keeping all pretrained components frozen, requiring no offline training or additional supervision. Extensive experiments across diverse combinations show that MatRes yields significant gains in both restoration and geometric alignment compared to using either restoration or matching models alone. MatRes offers a practical and widely applicable solution for real-world scenarios where users commonly capture multiple images of a scene with varying viewpoints and quality, effectively addressing the often-overlooked mutual interference between matching and restoration.

Problem

• image restoration
• geometric matching
• mutual interference
• viewpoint changes
• image degradation

Innovation

• zero-shot adaptation
• test-time adaptation
• image restoration
• geometric matching
• conditional similarity