Balanced Pedestrian Attribute Recognition for Improved Attribute-based Person Retrieval

Attribute-based person retrieval is a crucial task in surveillance systems since it enables the search for an individual if no image of the suspect is available. Searches can be based on soft biometric characteristics, understood as semantic attributes people typically use to describe each other. Examples are gender, age, or a description of the person's clothing and accessories. One possible solution is Pedestrian Attribute Recognition (PAR). Resulting attribute predictions are compared to the query attributes with the model's predictions. However, related literature indicates approaches tend to either perform well regarding the recognition of single attributes or achieve superior retrieval results at the expense of an increased number of false positives. This work aims to improve single-attribute recognition and attribute-based retrieval simultaneously by applying spatial projection and normalization modules. Spatial projection extracts features for different receptive fields to account for global as well as local attributes. Our PARNorm module normalizes attribute logits attribute-and instance-wise to enhance label-based PAR metrics without hurting instance-based performance. Furthermore, we propose a novel evaluation metric for attribute-based person retrieval that considers the degree of match between attribute queries and gallery images instead of using binary relevance labels. This allows for more realistic estimates of retrieval accuracy, which correlate strongly with the visual impression of the rankings. Our model outperforms the current state-of-the-art in generalization on the UPAR dataset and its challenge version.