Semantic Matching Based on Semantic Segmentation and Neighborhood Consensus

Establishing dense correspondences across semantically similar images is a challenging task, due to the large intra-class variation caused by unconstrained setting of images, which prone cause matching errors. To suppress potential ambiguity, NCNet explores neighborhood consensus pattern in 4D space all possible correspondences, based on assumption that correspondence continuous space. We retain constraint, while introducing semantic segmentation information into features, makes them more distinguishable and reduces ambiguity from feature perspective. Specifically, we combine network extract features convolution explore 4D-space context consistency. Experiments demonstrate our algorithm has good performances can improve accuracy.