Centroid Tuplet Loss for Person Re-Identification

Abstract

Person Re-Identification (Person Re-ID) is an important computer vision task in area surveillance, in which the goal is to match a person’s identity across different cameras or locations in videos or image sequences. To solve this task, Deep Metric Learning with the combination of different neural networks and metric losses such as Triplet Loss has become a common framework and achieved several remarkable results on benchmark datasets. However, Deep Metric Learning loss functions often depend on delicately sampling strategies for faster convergence and effective learning. These common sampling strategies usually rely on calculating embedding distances between samples in training datasets and selecting the most useful triplets or tuplets of images to consider, which makes these methods computationally expensive and may incur the risk of causing sample bias. Additionally, Triplet Loss also appears fragile to outliers and noisy labels. In this paper, we designed a centroid-based metric loss function, Centroid Tuplet Loss, which uses randomly selected mean centroid representations of classes in each mini-batch to achieve better retrieval performance. Experiments on two widely used Person Re-ID datasets, Market-1501 and CUHK03 dataset, demonstrates the effectiveness of our method over existing state-of-the-art methods.