Viewpoint robust knowledge distillation for accelerating vehicle re-identification

Abstract Vehicle re-identification is a challenging task that matches vehicle images captured by different cameras. Recent approaches exploit complex deep networks to learn viewpoint robust features for obtaining accurate results, which causes large computations in their testing phases restrict the speed. In this paper, we propose knowledge distillation (VRKD) method accelerating re-identification. The VRKD consists of teacher network and simple student network. Specifically, uses quadruple directional features. only contains shallow backbone sub-network global average pooling layer. distills from via minimizing Kullback-Leibler divergence between posterior probability distributions resulted networks. As result, speed significantly accelerated since small demanded. Experiments on VeRi776 VehicleID datasets show proposed outperforms many state-of-the-art with better performance.