Inter-Attribute awareness for pedestrian attribute recognition

作者：

Highlights：

• We propose an end-to-end inter-attribute awareness via vector-neuron capsules network for PAR (IAA-Caps). It explicitly considers the inter-attribute relations to increase the PAR accuracy.

• To the best of our knowledge, we are the first to utilize vector-neuron capsules for the PAR task. The CapsL is utilized to reflect the likelihood of existence of a pedestrian attribute, and the CapsO is used to build inter-attribute relations.

• We perform comprehensive experiments on several PAR datasets, including PETA, PA-100K, RAPv1, and RAPv2. Our IAA-Caps method out-performs existing methods and achieves state-of-the-art performance.

摘要

•We propose an end-to-end inter-attribute awareness via vector-neuron capsules network for PAR (IAA-Caps). It explicitly considers the inter-attribute relations to increase the PAR accuracy.•To the best of our knowledge, we are the first to utilize vector-neuron capsules for the PAR task. The CapsL is utilized to reflect the likelihood of existence of a pedestrian attribute, and the CapsO is used to build inter-attribute relations.•We perform comprehensive experiments on several PAR datasets, including PETA, PA-100K, RAPv1, and RAPv2. Our IAA-Caps method out-performs existing methods and achieves state-of-the-art performance.

论文关键词：Pedestrian attribute recognition,Inter-Attribute awareness,Vector-Neuron capsules

论文评审过程：Received 12 January 2022, Revised 11 May 2022, Accepted 16 June 2022, Available online 18 June 2022, Version of Record 22 June 2022.

论文官网地址：https://doi.org/10.1016/j.patcog.2022.108865