Cascade Network with Guided Loss and Hybrid Attention for Finding Good Correspondences

Finding good correspondences is a critical prerequisite in many feature based tasks. Given putative correspondence set of an image pair, we propose neural network which finds correct by binary-class classifier and estimates relative pose through classified correspondences. First, analyze that due to the imbalance number wrong correspondences, loss function has great impact on classification results. Thus, new Guided Loss can directly use evaluation criterion (Fn-measure) as guidance dynamically adjust objective during training. We theoretically prove perfect negative correlation between Fn-measure, so always trained towards direction increasing Fn-measure maximize it. then hybrid attention block extract feature, integrates Bayesian attentive context normalization (BACN) channel-wise (CA). BACN mine prior information better exploit global CA capture complex channel enhance awareness network. Finally, our block, cascade designed gradually optimize result for more superior performance. Experiments have shown achieves state-of-the-art performance benchmark datasets. Our code will be available https://github.com/wenbingtao/GLHA.