Learning Discriminative Context-Aware Keypoints Representations for Resolving Ambiguous Matches

Abstract

In the feature matching problem, local keypoint representations are often not sufficiently distinctive to disambiguate repetitive textures. State-of-the-art matching pipelines encode global information and embed context into keypoint descriptors to resolve this issue. In this thesis, we evaluate the failure modes of the state-ofthe- art method for image matching. We identify the problem that including global context to keypoint representations can sometimes eliminate their distinctiveness. We propose to enhance the learning of the state-of-the-art pipeline by adding a metric learning component to its objective function. By learning more distinctive global context-aware keypoint descriptors, we recover the filtered matches without the loss in matching precision.