A K-Nearest-Neighbor-Pooling method for graph matching

摘要

In the real-word matching, deformation, outliers and other emerging variations are inseparable conditions which make matching increasingly difficult. One way to solve this challenging problem is raising an effective graph matching method which is flexible to non-rigid objects. As a good structure representation, graph can nicely describe objects with meaningful information. Meanwhile, feature pooling plays a central role in feature sparsification. In this paper, we introduce a K-Nearest-Neighbor-Pooling Matching (KNNPM) method, which adopts feature pooling into the graph matching framework with lower solving complexity. The proposed algorithm evaluates each candidate matching with different weights on its k-nearest neighbors (KNN) by taking locality as well as sparsity into consideration. It also has wide range of applicability with generality in nowadays computer vision. In addition, our comparative and extensive experiments show the robustness and great improvements in matching accuracy.