Search:
Match:
2 results

Analysis

This paper addresses the critical problem of outlier robustness in feature point matching, a fundamental task in computer vision. The proposed LLHA-Net introduces a novel architecture with stage fusion, hierarchical extraction, and attention mechanisms to improve the accuracy and robustness of correspondence learning. The focus on outlier handling and the use of attention mechanisms to emphasize semantic information are key contributions. The evaluation on public datasets and comparison with state-of-the-art methods provide evidence of the method's effectiveness.
Reference

The paper proposes a Layer-by-Layer Hierarchical Attention Network (LLHA-Net) to enhance the precision of feature point matching by addressing the issue of outliers.

Analysis

This paper introduces SC-Net, a novel network for two-view correspondence learning. It addresses limitations of existing CNN-based methods by incorporating spatial and cross-channel context. The proposed modules (AFR, BFA, PAR) aim to improve position-awareness, robustness, and motion field refinement, leading to better performance in relative pose estimation and outlier removal. The availability of source code is a positive aspect.
Reference

SC-Net outperforms state-of-the-art methods in relative pose estimation and outlier removal tasks on YFCC100M and SUN3D datasets.