Inter-Attribute awareness for pedestrian attribute recognition

The task of pedestrian attribute recognition (PAR) is to distinguish a series of person semantic attributes. Generally, existing methods adopt multi-label classification algorithms to tackle the PAR task by utilizing multiple attribute labels. Despite remarkable progress, this kind of method normally ignores relations between different attributes. In order to be aware of relations between attributes, we propose an inter attribute aware network via vector-neuron capsule for PAR (IAA-Caps). Our IAA-Caps method replaces traditional one-dimensional scalar neurons with two-dimensional vector-neuron capsules by embedding them in IAA-Caps. Specifically, during IAA-Caps training, one dimension in capsules is used to recognize different attributes, and the other dimension is used to strengthen the relations of different attributes. Through considering inter-attribute relations, compared with previous methods that use a heavyweight backbone (e.g., ResNet50 or BN-Inception), a more lightweight backbone (i.e., OSNet) can be adopted in our proposed IAA-Caps to achieve better performance. Experiments are conducted on several PAR benchmark datasets, including PETA, PA-100K, RAPv1, and RAPv2, demonstrating the effectiveness of the proposed IAA-Caps. In addition, experiments also show that the proposed method can improve the performance of PAR on different backbones, showing its generalization ability.(c) 2022 Elsevier Ltd. All rights reserved.