ParaFormer: Parallel Attention Transformer for Efficient Feature Matching

Heavy computation is a bottleneck limiting deep-learning-based feature matching algorithms to be applied in many real-time applications. However, existing lightweight networks optimized for Euclidean data cannot address classical tasks, since sparse keypoint based descriptors are expected matched. This paper tackles this problem and proposes two concepts: 1) novel parallel attention model entitled ParaFormer 2) graph U-Net architecture with attentional pooling. First, fuses features positions through the concept of amplitude phase, integrates self- cross-attention manner which achieves win-win performance terms accuracy efficiency. Second, proposed pooling, ParaFormer-U variant significantly reduces computational complexity, minimize loss caused by downsampling. Sufficient experiments on various applications, including homography estimation, pose image matching, demonstrate that state-of-the-art while maintaining high The efficient comparable less than 50% FLOPs attention-based models.