S³-Match: Common-View Aligned Image Matching via Self-Supervised Keypoint Selection


Shizhen Li (Xi'an Jiaotong University), Jingcheng Liu (Xi'an Jiaotong University), Jianwu Fang (Xi'an Jiaotong University), DeZheng Gao (Xi'an Jiaotong University), Jianru Xue (Xi'an Jiaotong University)
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)Supplementary

Abstract

This paper introduces S³-Match, a common-view aligned image matching algorithm via self-supervised keypoint selection. The most common image matching methods depend on sparse interest points to minimize dependence on non-essential information and to effectively manage significant distortions, occlusions, or noise. Nonetheless, the repeatability of interest points and their reliable description often degrade in scenes with sparse textures or when there are changes in appearance due to varying viewpoints and lighting conditions. To overcome these challenges, S³-Match employs a quality score that autonomously identifies feature points with high distinctiveness and stability during the training phase. Furthermore, it incorporates a cross-attention mechanism that aligns features within the common-view areas across images. This alignment provides consistent feature information across images, and focuses subsequent self-supervised keypoint extraction and feature description on these common-view regions. Experimental results demonstrate that S³-Match significantly outperforms SuperPoint in terms of keypoint selection consistency and uniformity. It also exhibits superior performance in pose estimation tasks and surpasses other advanced algorithms in computational efficiency. Additionally, we have validated a variant of S³-Match that does not rely on cross-image information, capable of meeting a broader range of application needs.

Citation

@inproceedings{Li_2024_BMVC,
author    = {Shizhen Li and Jingcheng Liu and Jianwu Fang and DeZheng Gao and Jianru Xue},
title     = {S³-Match: Common-View Aligned Image Matching via Self-Supervised Keypoint Selection},
booktitle = {35th British Machine Vision Conference 2024, {BMVC} 2024, Glasgow, UK, November 25-28, 2024},
publisher = {BMVA},
year      = {2024},
url       = {https://papers.bmvc2024.org/0203.pdf}
}


Copyright © 2024 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection