Person Re-Identification via Generalized Class Prototypes

TL;DR

This paper tackles person Re-ID by moving beyond single-centroid or instance-based representations to a Generalized Class Prototype (GCP) framework that generates multiple prototypes per class via an autoregressive transformer decoder conditioned on a memory of gallery features. The approach explicitly models intra-class diversity and inter-class separation, using a loss that combines triplet supervision with a diversity regularizer to prevent prototype collapse. Experiments on CUHK03-NP, Market-1501, and MSMT17 show that GCP improves retrieval metrics over centroid- and instance-based baselines, with performance enhanced by stronger backbones such as TransReID and PHA. A key contribution is the ability to fix the number of prototypes per class ($N$) to tailor accuracy and mean average precision ($mAP$) to application needs, enabling practical, scalable Re-ID deployments.

Abstract

Advanced feature extraction methods have significantly contributed to enhancing the task of person re-identification. In addition, modifications to objective functions have been developed to further improve performance. Nonetheless, selecting better class representatives is an underexplored area of research that can also lead to advancements in re-identification performance. Although past works have experimented with using the centroid of a gallery image class during training, only a few have investigated alternative representations during the retrieval stage. In this paper, we demonstrate that these prior techniques yield suboptimal results in terms of re-identification metrics. To address the re-identification problem, we propose a generalized selection method that involves choosing representations that are not limited to class centroids. Our approach strikes a balance between accuracy and mean average precision, leading to improvements beyond the state of the art. For example, the actual number of representations per class can be adjusted to meet specific application requirements. We apply our methodology on top of multiple re-identification embeddings, and in all cases it substantially improves upon contemporary results

Figures (11)

• Figure 1: An overview of prototype-based Re-ID. The solid circles/squares denote the feature vectors of each image instance, while the empty circles represent query vectors. The two plots on the right depict the feature vector distribution in the embedding space. The upper-right plot illustrates instance-based Re-ID, which typically yields lower precision. The plot depicts a specific case where among the five nearest instances from the query, two are false positives. The lower-right plot demonstrates how our prototype-based Re-ID approach enhances precision by representing all instances from the same class using a distribution-aware feature vector.
• Figure 2: A comparison among (a) instance-based, (b) centroid-based, and (c) prototype-based Re-ID. The images enclosed by the colored rectangles are the query images and the small colored dots are the class representatives. The dashed lines indicate the distance between the query image and the class representation. In (a), (b), and (c), the query image is assigned to the class of the nearest instance, centroid, and prototype (our method) respectively.
• Figure 3: The proposed attention-based GCP model for person Re-ID. Small blue dots represent the extracted features $\mathcal{F}(x)$ from the backbone, while large green dots indicate the output prototypes from the model. PH denotes the prototype heads. For clarity, the figure displays only two of the generated prototypes.
• Figure 4: Prototype generation during inference via the proposed GCP model. Each ellipse signifies a feature vector, with colors (e.g., red, green, blue) indicating the associated camera ID. When dealing with a particular query captured by a green camera, feature vectors marked in green are excluded during prototype generation.
• Figure 5: Top-$k$ accuracy on the (a) CUHK03-NP, (b) Market-1501, and (c) MSMT17 datasets.