A Transformer based Multi task Model for Attribute based Person Retrieval

Person retrieval is a crucial task in video surveillance. While searching for persons-of-interest based on so-called query images gains much interest in the research community, attribute-based approaches are rarely studied. Attribute-based person retrieval takes a person’s semantic attributes as input and provides a ranked list of search results that match the description. Typically, such approaches either build on a pedestrian attribute recognition approach or learn a joint feature space between attribute descriptions and image data. In this work, both approaches are combined in a multi-task model to benefit from the advantages of both procedures. Moreover, transformer modules are incorporated to increase performance further. Experimental evaluation proves the effectiveness of the approach and shows that the proposed architecture outperforms the baselines significantly.