Structural Relation Enhancement for Open-Vocabulary Pedestrian Attribute Recognition

Open-Vocabulary Pedestrian Attribute Recognition (OV-PAR) aims to recognize and understand novel attributes that do not appear in the training set. Although much progress has been made in existing open-vocabulary recognition methods for natural scenes, there are two major challenges in fine-grained attribute recognition of pedestrian images. (1) Intra-group pedestrian attributes often lie in local regions with minimal semantic differences (e.g., sports and casual shoes) and are thus difficult to distinguish. (2) Inter-group attributes exhibit significantly different semantic hierarchies (e.g., action vs. shoes), requiring the model to balance global understanding and fine-grained discernment. To address the above challenges, we propose a Structural Relation Enhancement Network (SRENet) that jointly models spatial, semantic, and attribute structural relations to capture local discriminative details, global contextual information, and common knowledge among intra-group attributes, thereby enhancing generalization ability for OV-PAR. Specifically, we perform attribute-fair spatial-semantic alignment and multi-level semantic fusion to distinguish subtle attribute differences while maintaining global context understanding, thereby preventing overfitting to particular attribute types. Moreover, we propose an intra-group knowledge sharing sentence retrieval paradigm that allows novel attributes to inherit base attribute representations, thereby improving generalization to novel attributes. This work presents a systematic study of OV-PAR and introduces two benchmarks, evaluating twenty representative methods. Extensive experiments demonstrate that our SRENet significantly outperforms state-of-the-art methods. The code and benchmark will be released on https://github.com/Event-AHU/OpenPAR.