Distributed Implementation for Person Re-Identification

Abstract

Person re-identification is to associate people across different camera views at different locations and time. Current computer vision algorithms on person re-identification mainly focus on performance, making it unsuitable for distributed systems. For distributed system, computational complexity, network usage, energy consumption and memory requirement are as important as the performance. In this paper, we compare the merits of the current algorithms. We consider three key algorithms Keep It Simple and Straightforward MEtric (KISSME), Symmetry-Driven Accumulation of Local Features (SDALF) and Unsupervised Saliency Matching (USM). The advantage of SDALF, and USM is that they are unsupervised methods so training is not required but computationally many time expensive than KISSME. Saliency based method is superior in performance but also has the longest feature length. As the feature needs to be transmitted from one camera to other in distributed system, this mean higher energy consumption and longer time delay. Among these three, KISSME offers a balance between performance, complexity and feature lengths hence more suitable for distributed systems.