ACMatch: Improving context capture for two-view correspondence learning via adaptive convolution

Abstract

Two-view correspondence learning plays a pivotal role in the field of computer vision. However, this task is beset with great challenges stemming from the significant imbalance between true and false correspondences. Recent approaches have started leveraging the inherent filtering properties of convolution to eliminate false matches. Nevertheless, these methods tend to apply convolution in an ad hoc manner without careful design, thereby inheriting the limitations of convolution and hindering performance improvement. In this paper, we propose a novel convolution-based method called ACMatch, which aims to meticulously design convolutional filters to mitigate the shortcomings of convolution and enhance its effectiveness. Specifically, to address the limitation of existing convolutional filters of struggling to effectively capture global information due to the limited receptive field, we introduce a strategy to help them obtain relatively global information by guiding grid points to incorporate more contextual information, thus enabling a global perspective for two-view learning. Furthermore, we recognize that in the context of feature matching, inliers and outliers provide fundamentally different information. Hence, we design an adaptive weighted convolution module that allows the filters to focus more on inliers while ignoring outliers. Extensive experiments across various visual tasks demonstrate the effectiveness, superiority, and generalization. Notably, ACMatch attains an AUC@5° of 35.93% on YFCC100M without RANSAC, surpassing the previous state-of-the-art by 5.85 absolute percentage points and exceeding the 35% AUC@5° bar for the first time. Our code is publicly available at https://github.com/ShineFox/ACMatch.