Learning Deep Neural Networks for Vehicle Re-ID with Visual-spatio-temporal Path Proposals

摘要

Vehicle re-identification is an important problem and has many applicationsin video surveillance and intelligent transportation. It gains increasingattention because of the recent advances of person re-identificationtechniques. However, unlike person re-identification, the visual differencesbetween pairs of vehicle images are usually subtle and even challenging forhumans to distinguish. Incorporating additional spatio-temporal information isvital for solving the challenging re-identification task. Existing vehiclere-identification methods ignored or used over-simplified models for thespatio-temporal relations between vehicle images. In this paper, we propose atwo-stage framework that incorporates complex spatio-temporal information foreffectively regularizing the re-identification results. Given a pair of vehicleimages with their spatio-temporal information, a candidatevisual-spatio-temporal path is first generated by a chain MRF model with adeeply learned potential function, where each visual-spatio-temporal statecorresponds to an actual vehicle image with its spatio-temporal information. ASiamese-CNN+Path-LSTM model takes the candidate path as well as the pairwisequeries to generate their similarity score. Extensive experiments and analysisshow the effectiveness of our proposed method and individual components.