CANOPY:Cross-view Attribute recogNition via OPen vocabularY

Pedestrian Attribute Recognition (PAR) predicts semantic attributes such as clothing, gender, carried objects, and accessories from pedestrian images. Although recent PAR models achieve strong performance on ground-view surveillance, they rely on a fixed attribute vocabulary, making them difficult to generalize across UAV benchmarks with heterogeneous annotations[1]. Existing datasets define different semantic spaces, causing classifier heads to become dataset-specific and restricting cross-dataset evaluation to only overlapping attributes. We propose CANOPY, a language-aligned PAR framework that replaces fixed classifiers with open-vocabulary semantic prediction. By representing each attribute through natural-language embeddings and aligning them with visual features, CANOPY enables a single model to recognize heterogeneous attribute vocabularies while improving robust cross-view generalization.