Multi-expert learning for fusion of pedestrian detection bounding box

Abstract

Performance of pedestrian detection, which is one of the essential tasks in automatic drive, relies heavily on a large number of labels. Some researchers proposed unsupervised domain adaptive frameworks to improve the detection accuracy in wild datasets to reduce the need for labels. However, it is not a down-to-earth and cost-effective way for deploying these frameworks in practical engineering because it needs both source and target data for training. Unlike the former research, this work presents a new fine-tuning method without using source and target data for unsupervised detection. In this work, different well-trained models from the source domain are regarded as less-accurate experts in the wild domain, where a multi-expert learning algorithm is applied to learn from the difference between these models and fuse bounding boxes to present more accurate detection results. Experimental results on three common pedestrian detection datasets show that our method can efficiently improve the detection accuracy under unsupervised settings. Our method can also achieve better performance without source and target data involved compared with state-of-the-art works.