English

$\mathbb{X}$Resolution Correspondence Networks

Computer Vision and Pattern Recognition 2021-03-25 v2

Abstract

In this paper, we aim at establishing accurate dense correspondences between a pair of images with overlapping field of view under challenging illumination variation, viewpoint changes, and style differences. Through an extensive ablation study of the state-of-the-art correspondence networks, we surprisingly discovered that the widely adopted 4D correlation tensor and its related learning and processing modules could be de-parameterised and removed from training with merely a minor impact over the final matching accuracy. Disabling these computational expensive modules dramatically speeds up the training procedure and allows to use 4 times bigger batch size, which in turn compensates for the accuracy drop. Together with a multi-GPU inference stage, our method facilitates the systematic investigation of the relationship between matching accuracy and up-sampling resolution of the native testing images from 1280 to 4K. This leads to discovery of the existence of an optimal resolution X\mathbb{X} that produces accurate matching performance surpassing the state-of-the-art methods particularly over the lower error band on public benchmarks for the proposed network.

Keywords

Cite

@article{arxiv.2012.09842,
  title  = {$\mathbb{X}$Resolution Correspondence Networks},
  author = {Georgi Tinchev and Shuda Li and Kai Han and David Mitchell and Rigas Kouskouridas},
  journal= {arXiv preprint arXiv:2012.09842},
  year   = {2021}
}

Comments

Preprint. Code is available at https://xyz-r-d.github.io/xrcnet

R2 v1 2026-06-23T21:03:34.038Z