Augmented inputs for surveillance re-identification

Abstract

Person Re-Identification (Re-ID) is one of the important applications for Surveillance. However, the performance of Re-ID is dependent on the input quality, which cannot be guaranteed from the surveillance systems. We explored the technique from Gait Re-ID to address viewpoint changes. From our findings, adding horizontally mirrored image into an auxiliary pipeline can achieve a modest performance uplift in our test (0.8% net increase in mean average precision and 0.9% increase in Rank-1 accuracy) in MARS dataset. This extra pipeline can be substituted by Heterogeneous Input Triplet Loss (hiTri) for minimal performance loss. The overall performance of the proposed method outperforms state-of-the-art techniques on well-known datasets. Further investigations on other auxiliary input types are warranted.