Attentive Part-Based Alignment Network for Vehicle Re-Identification

Abstract

Vehicle Re-identification (Re-ID) has become a research hotspot along with the rapid development of video surveillance. Attention mechanisms are utilized in vehicle Re-ID networks but often miss the attention alignment across views. In this paper, we propose a novel Attentive Part-based Alignment Network (APANet) to learn robust, diverse, and discriminative features for vehicle Re-ID. To be specific, in order to enhance the discrimination of part features, two part-level alignment mechanisms are proposed in APANet, consisting of Part-level Orthogonality Loss (POL) and Part-level Attention Alignment Loss (PAAL). Furthermore, POL aims to maximize the diversity of part features via an orthogonal penalty among parts whilst PAAL learns view-invariant features by means of realizing attention alignment in a part-level fashion. Moreover, we propose a Multi-receptive-field Attention (MA) module to adopt an efficient and cost-effective pyramid structure. The pyramid structure is capable of employing more fine-grained and heterogeneous-scale spatial attention information through multi-receptive-field streams. In addition, the improved TriHard loss and Inter-group Feature Centroid Loss (IFCL) function are utilized to optimize both the inter-group and intra-group distance. Extensive experiments demonstrate the superiority of our model over multiple existing state-of-the-art approaches on two popular vehicle Re-ID benchmarks.