Large margin metric learning for multi-view vehicle re-identification

Abstract

In this paper we propose a novel large margin metric learning scheme for vehicle re-identification (re-ID), which aims to maximize the distance margins among different vehicles in the embedding feature space. Inspired by the fact that in the embedding feature space vehicles with the same identity are usually scattered sparsely due to vehicles’ multi-view appearance, our method is designed to make samples with the same identities more compact and meanwhile project instances with different categories to separate regions in the embedding space. To make training process more efficient, in the first stage, only the Softmax Loss is adopted, and in the second stage, by computing classification hyperplanes between different vehicle identities, a large margin loss is defined to maximize the distances between training samples and their corresponding hyperplanes. Besides, a new sampling method is adopted to find the hardest samples which are close to their hyperplanes and a kernelized re-ranking method is applied to further boost the performance. Compared with state-of-the-art approaches, our method achieves superior results with efficient training and inference process and only the identity as supervision signal. Experimental results on three most popular datasets show that our system produce promising results, and notably on the VeRi-776 dataset, our method can reach the best record with Rank1 96.81% and mAP 80.95%.