Boosting Generalization Performance in Person Re-identification

Abstract

Generalizable person re-identification (ReID) has gained significant attention in recent years as it poses greater challenges in recognizing individuals across different domains and unseen scenarios. Existing methods are typically limited to a single visual modality, making it challenging to capture rich semantic information across different domains. Recently, pre-trained vision-language models like CLIP have shown promising performances in various tasks by linking visual representations with their corresponding text descriptions. This enables them to capture diverse high-level semantics from the accompanying text and obtain transferable features. However, the adoption of CLIP has been hindered in person ReID due to the labels being typically index-based rather than descriptive texts. To address this limitation, we propose a novel Cross-modal framework wIth Conditional Prompt (CICP) framework based on CLIP involving the Description Prompt Module (DPM) that pre-trains a set of prompts to tackle the lack of textual information in person ReID. In addition, we further propose the Prompt Generalization Module (PGM) incorporates a lightweight network that generates a conditional token for each image. This module shifts the focus from being limited to a class set to being specific to each input instance, thereby enhancing domain generalization capability for the entire task. Through extensive experiments, we show that our proposed method outperforms state-of-the-art (SOTA) approaches on popular benchmark datasets.

The work was supported in part by the National Natural Science Foundation of China under Grant 82172033, U19B2031, 61971369, 52105126, 82272071, 62271430, and the Fundamental Research Funds for the Central Universities 20720230104.