Enhancing Person Retrieval with Joint Person Detection, Attribute Learning, and Identification

摘要

Person re-identification receives increasing attention in recent years. However, most works assume the persons have been well cropped from the whole scene images, and only focus on learning features and metrics. This paper considers the person re-identification problem in a real-world scenario, which should consider detection and identification simultaneously. This paper proposes a multi-task learning framework for person retrieval in the wild. Person attribute learning is exploited in our framework to enhance person retrieval. Our work consists of two main contributions: (1) we present a 11 image-level attribute annotations for each image in the large-scale PRW [27] dataset, and (2) we develop an end-to-end person retrieval framework which jointly learns person detector, attribute detectors, and visual embeddings in a multi-task learning manner. We evaluate the effectiveness of the proposed approach on two tasks, i.e. person attribute recognition and person re-identification. Experimental results have demonstrated the effectiveness of the proposed approach.