Local Feature Matching with Wavelet Attention Transformer

Learning-based feature description algorithms have received widespread attention due to the rapid development of convolutional neural networks and their excellent performance in feature-matching tasks. In this paper, we propose a detector-free multi-scale feature aggregation matching network based on a wavelet transformer to improve the discrimination ability and robustness of learning feature descriptors. Different from the method of directly using the cost module to obtain matching, the global receptive field provided in our method allows for generating dense matching in low-texture regions. Experimental results on the matching task show that our feature description algorithm is progressive in actual visual tasks.