Self-aware Multi-inductor Net: Cross Domain Person Re-identification Based on Global Self-attention and Multi-granularity Supervision

One of the biggest challenges on deep-learning-based person re-identification technology is its serious degradation of cross-domain testing performance. The poor cross-domain generalization could attribute to the deficiency on extracting typical and adaptive feature representations from pedestrian images. The pure convolutional baseline has two major defects: for one thing, its weights are relatively rigid due to its static architecture without enough dynamically selective mechanism; for another, the guidance to tune learnable parameters is insufficient because there is only one way signal for supervision. Therefore, it hinders networks to adapt inputs flexibly and fails to induce networks to utmost mine the potential information hidden in the human body structure. In this paper, a convolutional-based neural network with global self-attention and multi-granularity supervision is proposed to improve the adaptiveness of feature extraction, named as “Self-aware Multi-inductor Net (SMNet)”. It only trains on source domain in a supervised manner, but could improve the generalization performance on the target domain by a nice bit. Experiment results on classic public datasets show that our proposed method could effectively upgrade the cross-domain generalization for person re-identification tasks.