Beyond outlier removal: Integrated ensemble matching for accurate image keypoint correspondence

Code

• github.com/ShowStopperTheSecond/MatchBeyondOutliers
  none★ 0

Abstract

This paper presents a novel two-tier matching approach for robust feature correspondence and geometric transformation estimation in computer vision tasks. Our method combines sub-descriptor matching and multi-descriptor ensembling to significantly improve the accuracy and reliability of keypoint correspondences across images. Our approach is evaluated on both classic hand-crafted algorithms and deep learning-based methods using the HPatches, IMC2022, and YFCC100M datasets. We use a pretrained network and enhance it to incorporate additional descriptor heads optimized for our matching strategy. The two-tier matching strategy enables direct geometric transformation estimation without separate outlier removal in many cases, potentially streamlining computer vision pipelines. Results demonstrate that our approach substantially outperforms traditional single-descriptor methods, achieving over 95% correct matches for classic algorithm combinations and up to 96% for our enhanced learning-based approaches. Our approach can establish reliable correspondences without making any assumption about the mathematical relation between the matches. Additionally, we explore applications of our method in scene matching.

Reproductions