TCDesc: Learning Topology Consistent Descriptors for Image Matching

The triplet loss is widely used in learning the local descriptors for image matching. However, existing loss-based methods, like HardNet and DSM, employ point-to-point distance metric, which neglects neighborhood information of descriptors. Considering fact that structures matching would be similar under ideal condition, this paper aims to learn topology-consistent (TCDesc). To end, we first propose linear combination weight as topology depict each descriptor, where difference between center descriptor its neighbors minimized. For global comparison, then define a vector by using weights. Next, beyond Euclidean distance, with vectors indicate topological Furthermore, an adaptive weighting strategy jointly minimize loss. Experimental results on four widely-used datasets, i.e., UBC PhotoTourism, HPatches, W1BS Oxford, demonstrate our method can effectively improve performance both DSM.