Body Part-Based Person Re-identification Integrating Semantic Attributes

摘要

Person re-identification is an important task to match pedestrian images from disjointed camera views. Most existing approaches seek a target by learning a global ranking model with the entire information of the image, which ignores the individuality of each person and the different contributions of each body part. In this work, considering the different contributions of each pedestrian’s feature representation and each body part, Sample-Specific SVM (SSSVM) is learned for each individual upper and lower body part. Finally, we combine the distances of the upper and lower body parts between images to obtain their final distances. In addition, we introduce semantic attributes as the complement to appearance features, thus making the feature representation robust to viewpoint and illumination changes. Experimental results on three challenging datasets (including VIPeR, QMUL GRID and PRID 2011) demonstrate that the personal re-identification algorithm proposed in this work performs favourably against the state-of-the-art approaches.