Improved Nearest Neighbor Distance Ratio for Matching Local Image Descriptors

This paper presents a novel matching strategy, called Improved Nearest Neighbor Distance Ratio, for matching local image descriptors. We first empirically analyze to what extent correspondences underlie the second nearest neighbor or even the third and so on. Based on the solid analysis, we propose to improve the widely-used Nearest Neighbor Distance Ratio (NNDR) by matching local descriptors not only based on the first nearest neighbor, but also making use of the second nearest neighbor appropriately. The proposed INNDR is evaluated against NNDR on a set of benchmark datasets. Our experiments will demonstrate that INNDR generally outperforms the traditional NNDR in both matching accuracy and recall vs 1-precision.