Probabilistic Distribution Alignment for Text-Based Person Retrieval

Text-based person retrieval is a cross-modal task that seeks to match pedestrian images with their corresponding textual descriptions. A key challenge in this task arises from the inherent one-to-many relationships: a single image can correspond to multiple descriptions, and a single description may relate to several images. Conventional deterministic embedding methods, which map images and texts to fixed feature vectors, struggle to capture such complex relationships effectively. To overcome this limitation, we introduce Probabilistic Distribution Alignment (PDA), a framework that represents both pedestrian images and text as probabilistic distributions and models the interactions between visual and linguistic modalities. PDA comprises three main components. First, Distributional Representation Modeling (DRM) encodes images and text into Gaussian distributions using a specially designed distance metric, allowing the model to capture uncertainty in the representations. Second, Cross-Modal Containment (CMC) aligns the distributions of text and masked text with their associated image distributions to strengthen semantic correspondence. Third, Intra-Modal Containment (IMC) enforces structured learning within each modality by embedding distributions alongside their masked variants, improving robustness to incomplete observations. Experiments on standard benchmarks demonstrate that PDA achieves superior performance compared with state-of-the-art methods, effectively handling ambiguity and cross-modal variability. These results highlight probabilistic distribution modeling as a powerful paradigm for vision-language alignment in pedestrian retrieval.