AAPMatcher: Adaptive attention pruning matcher for accurate local feature matching

Local feature matching, which seeks to establish correspondences between two images, serves as a fundamental component in numerous computer vision applications, such as camera tracking and 3D mapping. Recently, Transformer has demonstrated remarkable capability in modeling accurate correspondences for the two input sequences owing to its long-range context integration capability. Whereas, indiscriminate modeling in traditional transformers inevitably introduces noise and includes irrelevant information which can degrade the quality of feature representations. Towards this end, we introduce an adaptive attention pruning matcher for accurate local feature matching (AAPMatcher), which is designed for robust and accurate local feature matching. We overhaul the traditional uniform feature extraction for sequences by introducing the adaptive pruned transformer (APFormer), which adaptively retains the most profitable attention values for feature consolidation, enabling the network to obtain more useful feature information while filtering out useless information. Moreover, considering the fixed combination of self- and cross-APFormer greatly limits the flexibility of the network, we propose a two-stage adaptive hybrid attention strategy (AHAS), which achieves the optimal combination for APFormers in a coarse to fine manner. Benefiting from the clean feature representations and the optimal combination of APFormers, AAPMatcher exceeds the state-of-the-art approaches over multiple benchmarks, including pose estimation, homography estimation, and visual localization.