Outlier Removal Based on Motion Filtering and Adjustment

The image point correspondences established by off-the-shelf feature extractors usually contain a large number of outliers, which severely affects the effectiveness of feature matching and the performance of downstream tasks reliant on the matching results. Several recently proposed outlier removal methods leverage the motion consistency of correspondences by estimating a motion field and employ convolutional neural network (CNN) to reduce contamination from outliers to capture context. However, CNN inherently suffers from limitations in capturing global context, as the fixed and localized nature of their receptive fields makes it difficult for models to adaptively integrate long-range information, thereby constraining the performance of related methods. Departing from these methods that directly estimate motion fields using CNN, this paper explores estimating a high-quality motion field without reliance on CNN. To this end, a motion filtering and adjustment network (MFANet) is proposed to mitigate the impact of outliers during context capture. Specifically, a motion-filtering block is first designed to iteratively remove outliers and capture contextual information. Then, a regularization and adjustment block is designed to estimate an initial motion field, which is then refined for greater accuracy by incorporating additional positional information. The performance of MFANet is evaluated on both indoor and outdoor datasets for the tasks of outlier removal and relative pose estimation. Experimental results demonstrate that MFANet achieves superior performance compared to several existing methods.