ECI: Effective Channel Interaction for Person Search

Person search, a combination of person detection and person re-identification (re-ID), is one of the research hotspots of intelligent security. Person detection treats all people as one of a kind and focuses on the commonalities of people. In contrast, re-ID treats persons with different identities as different classes, focusing on the uniqueness of each individual. In order to better balance these two conflicting tasks, this paper proposes a person search network with effective channel interaction (ECI) to reduce the impact on model performance. First, the channel attention module ECA is inserted into the residual block of the backbone network ResNet50 using NAE as the baseline. The correlation between each channel and its immediate neighbours is strengthened with the help of 1D convolution. This channel interaction approach allows the model to learn autonomously which features of the input image are more dominant and thus adjust the feature weights appropriately. Next, the accuracy of the model is tested for each of the five cases of 1D convolutional kernel size and compare to the original model without the use of channel attention. Experimental results show that our method achieves the highest accuracy of 91.29% and 92.55% for mAP and top-1, respectively, in the large public dataset CUHK-SYSU.