GlobalAP: Global average precision optimization for person re-identification

Abstract

Average Precision (AP) measures the overall performance on the Person Re-Identification (ReID) task. Optimizing AP using all instances in the training set is accordingly an excellent choice for learning a discriminative ReID model. However, exploiting this method directly is unacceptable in practice due to the high cost of computation on the entire dataset. To this end, this paper proposes an effective and easy-to-use approach called GlobalAP that optimizes AP globally at negligible computational cost. More specifically, GlobalAP adopts a memory module to acquire the embedding features of all instances in the training set. To reduce the required computational complexity, GlobalAP utilizes only a few instances with high similarities to the query, to compute AP; this is because we observe that only these instances significantly affect AP and model optimization. Moreover, we propose to gradually increase the difficulty of GlobalAP to further encourage intra-class compactness and inter-class separability. Ultimately, GlobalAP can globally optimize AP and dramatically boost the model performance at negligible computational cost. We evaluate GlobalAP on six large-scale ReID datasets. Experimental results show that GlobalAP exhibits obvious advantages in terms of both computational efficiency and ReID accuracy.