Spatial and Channel Dual Significance Mining Algorithm for Object Re-identification

Existing object re-identification methods usually use global feature pooling layers to aggregate feature maps extracted via deep backbone networks to obtain the final image features. However, global feature pooling layers ignore the spatial and channel significance of feature maps, restricting the discriminative ability of the final image features. For that, in this paper, a spatial channel dual significance mining (SC-DSM) module is designed to comprehensively exploit the feature map’s significance from both spatial and channel dimensions to enhance the feature’s discriminative ability, thereby improving the object re-identification accuracy. The SC-DSM module consists of a spatial significance mining (SSM) sub-module and a channel significance mining (CSM) sub-module. On feature maps, the SSM sub-module builds spatial graphs to aggregate spatial neighbor nodes’ features and learn weights to exploit the spatial significance, while the CSM sub-module constructs channel graphs to aggregate channel neighbor nodes’ features and learn weights to mine the channel significance. Experimental results show that proposed method can outperform existing object re-identification methods on the most popular vehicle re-identification dataset, namely, VeRi776 and person re-identification dataset, namely, Market-1501.