On the Effect of Data Imbalance for Multi-Label Pedestrian Attribute Recognition

Abstract

Pedestrian attribute recognition has many applications in surveillance and attribute based query, tracking, and person re-identification. The recent trend in deep-learning based pedestrian attribute recognition is to use a shared CNN backbone for feature extraction and multiple subsequent branches for the individual branches. While this allows the end-to-end learning to simultaneously recognize multiple attributes, the data imbalance problem of most attributes becomes a challenge that has not been studied sufficiently for this application. This paper presents studies on how the cost adjustment method affects several common evaluation metrics. We also propose a two-stage training procedure, where an additional fine-tuning stage on the classifier layers only with class-balanced data is shown to improve recognition performances.