LRHW-AP: Using ranking-based metric as loss for Person Re-Identification

Abstract

Optimizing a ranking-based metric as the loss function, such as Average Precision (AP), has been found very effective in image retrieval tasks, but it has received less attention in Person Re-Identification (Re-ID). In this paper, Low Rank High Weight (LRHW) AP is proposed to apply the AP-optimizing method on the Re-ID task. LRHW-AP employs high weight on the low rank positive instances, which provides more information for model optimization than high rank positive instances and distribute in high gradient area. We propose a new pooling method called Power Activation Weighted Mean (PAWM) pooling which can unify a set of pooling methods because of a changeable activation function and a trainable parameter. Thus one can adjust and train PAWM to adapt to the target task to improve the model performance. Besides, we incorporate Warmup and Exponentially Decay Scheduler with a delay period, called Warmup Delay Exponentially Decay Scheduler, which brings further improvement. Through an extensive set of ablation studies, we verify that all methods mentioned above contribute to the performance boosts on Re-ID and the model achieves 95.3% rank-1 and 88.4% mAP on Market1501 with ResNet50.