Enhancing person re‐identification by late fusion of low‐, mid‐ and high‐level features

Person re‐identification is the process of finding people across different cameras. In this process, focus often lies in developing strong feature descriptors or a robust metric learning algorithm. While the two aspects are the most important steps in order to secure a high performance, a less explored aspect is late fusion of complementary features. For this purpose, this study proposes a late fusing scheme that, based on an experimental analysis, combines three systems that focus on extracting features and provide supervised learning on different abstraction levels. To analyse the behaviour of the proposed system, both rank aggregation and score‐level fusion are applied. The authors’ proposed fusion scheme increases results on both small and large datasets. Experimental results on VIPeR show accuracies 5.43% higher than related systems, while results on PRID450S and CUHK01 increase state‐of‐the‐art results by 10.94 and 14.84%, respectively. Furthermore, a cross‐dataset test shows an increased rank‐1 accuracy of 28.26% when training on CUHK02 and testing on VIPeR. Finally, an analysis of the late fusion shows aggregation to be better when individual results are unequally distributed within top‐10 while score‐level fusion provides better results when two individual results lie within top‐5 while the last lies outside top‐10.