Affine-invariant description of keypoint bundles for detecting partial near-duplicates in random images

A method for partial near-duplicate retrieval in random images is proposed and evaluated. Unlike the majority of existing methods, it is based on matching individual keypoints only (i.e. no analysis/verification of configuration constraints). The proposed description of keypoints incorporates affine-invariant representation of keypoint bundles (photometric and geometric properties of neighboring keypoints). The representation is flexible enough to accept a wide range of distortions. It is shown, using a small dataset of diversified images, that the method provides reasonably high performances at very low computational costs. The method can also be used as a highly-selective scheme of keypoint matching; its outputs often better reflect the “human-centric” concept of keypoints than typical approaches to keypoint matching.