Enhancing Text-Based Person Retrieval via Loss Balancing and Vision-Language Models

This paper presents the ZhCheng team's solution to the WWW 2025 challenge on Text-based Person Anomaly Search (TPAS). Text-based person retrieval aims to retrieve specific individuals using natural language descriptions. A key challenge in this field is recognizing anomalous behaviors in real-world scenarios, such as unusual actions, rather than common ones like walking or standing. Existing state-of-the-art methods underperform on this new task. After experiments, we identified two aspects-loss functions and architecture-that can significantly enhance performance. First, we introduce a strategy that leverages homoscedastic uncertainty to balance multiple loss functions. Second, considering the remarkable success of pretrained vision-language models (VLMs) in vision language tasks, we adopt a pretrained VLM as the backbone and fine-tune it on the existing benchmark. Experimental results demonstrate that our proposed approach effectively improves the performance of current baselines on this challenging task.