CSFormer: Cross-Scale Transformer for Feature Matching
Yuhang Zhou, Xi Cheng, Xiaomeng Zhai, Xue Lei, Сонглин Ду · 2023
Existing feature matching methods tend to use Transformer to extract the features of the image for image matching. However, Transformer pays more attention on global feature, rich local features of the image are not effectively used. Local features are usually unique, stable, distinguishable and repeatable, so adopting excellent local features can effectively match similar areas in two images. In this paper, a cross-scale image local feature extraction network is proposed, which extracts richer local features from the output of convolutional neural networks with different receptive fields, and reorganizes channel features at the semantic perception level. Based on extracting global features from the Transformer, we introduce deeper local features to construct a new image matching network with fused features. This structure enhances the richness of image features and makes up for the lack of local structure information. Extensive experiments on indoor and outdoor pose estimation results demonstrate that our CSFormer outperforms other existing state-of-the-art learned approaches by a large margin.