A Local Self-Similarity-Based Vehicle Detection Approach Using Single Query Image

Abstract

We present a generic vehicle detection approach using a single query image of vehicle to find similar objects in the test image. The proposed method is without any prior training or segmentation of the test image. The approach is based on computing local self-similarity (LSS) descriptors from query and test images that capture local internal geometric layout within the image. Descriptors from query and test image are matched as two-stage process in sliding window framework. In order to exploit usefulness of LSS for generic object detection, we make following contributions: (i) we present few novel ideas to discard the non-informative descriptors to reduce computational expense in feature matching. (ii) We propose a deformation tolerant version of sliding window-based matching framework rather than point-to-point matching. (iii) We also show that selection of landmark points from the query not only makes the algorithm faster but also improves the performance of detection. We evaluate our results on UIUC car dataset, and results clearly outperform earlier training free methods with 91% accuracy.