One-Shot Vehicle Re-Identification

Abstract

Although person re-identification has amassed great attention due to being a major part of intelligent video surveillance systems, however, there is another, although less seen, but important end of the intelligent video surveillance system, which is Vehicle re-identification. Vehicle re-identification’s importance is not just limited to video surveillance as it can be used as a part of intelligent traffic monitoring systems and can play a key role in forensic analysis. This thesis introduces “One-shot” methodology take on vehicle re-identification using hard-batch, offline, triplet loss function using "Resnet50" as the backbone for feature extraction. “One-shot” takes only one camera per vehicle to mine its way into the database. Triplet loss is a refined version of Siamese network. Where a simple Siamese network either take positive anchor (similar) or negative anchor (different) image of a vehicle while triplet-loss takes both positive and negative anchors in a single training point. All the computations are performed using mixed precision; allows speeding up computations, and loading larger network and batch size