Partner learning: A comprehensive knowledge transfer for vehicle re-identification

Abstract

The intra-class variability and inter-class similarity challenges caused by diverse viewpoints, illumination, and similar appearances are crucial in Re-Identification (Re-ID). Previous vehicle Re-ID methods propose to mine more discriminate and fine-grained clues for alleviating the problem, which costs extra computation and time during inference since the use of additional modules, e.g., detection modules, segmentation modules, or attention modules. We propose a multi-branch architecture to mining the discriminative and fine-grained information without additional time and computation cost during inference. Specifically, we focus on three problems: 1) how can knowledge transfer among multi-branches; 2) what knowledge should be utilized for more effective and more functional transfer; 3) where can be used as the input of multi-branches? For the first problem, we introduce a novel complementary learning scheme named partner learning which transfers the knowledge between global and local branches, and thus we only need the global branch during inference. For the second problem, we propose a hierarchical structural knowledge transfer (HSKT) approach to mine knowledge from partners in three different levels hierarchically. For the last problem, to effectively mine more fine-grained clues, we propose two local specifications: one supervised with the specification of the window area being discriminatively crucial as an expert knowledge while the other unsupervised with horizontal stripe cuts. Extensive ablation studies and experimental result discussions show the effectiveness of the proposed method.