Enhancing Person Re-identification Using Polynomial Expansion of Cross Entropy Loss

In the advance of machine learning, person re-identification (re-ID) algorithms have gained a dramatic improvement to identify a person without a clear face or frontal image in the real world. Since recent studies have found that the polynomial expansion of cross entropy loss function can learn more effectively than the original version on training neural networks for object detection tasks, we are motivated to utilize this finding to make an improvement on the deep metric learning for person re-ID. In this work, we utilize a linear combination of a polynomial cross entropy and a triplet loss function to train the well-known AGW baseline. Experimental results have shown that the proposed method outperforms the original AGW, reaching rank-1 accuracy of 96.4% (with mAP: 94.6) and rank-1 accuracy of 93.6% (with mAP: 91.4) on Market1501 and DukeMTMC datasets, respectively.