Dissecting Person Re-identification from the Viewpoint of Viewpoint

  • 2018-12-05 18:56:10
  • Xiaoxiao Sun, Liang Zheng
  • 1

Abstract

In person re-identification (re-ID),In person re-identification (re-ID), weusually refer the challenges of this task to variances in visual factors suchas the viewpoint, pose, illumination and background. In spite of acknowledgingthese factors to be influential, quantitative studies on how they affect are-ID system are still lacking.To gain insights in this scientific campaign,this paper makes an early attempt in studying a particular factor, viewpoint.We narrow the viewpoint problem down to the pedestrian rotation angle to obtainfocused conclusions. In this regard, this paper makes two contributions to thecommunity. First, we introduce a large-scale synthetic data engine, PersonX.Composed of hand-crafted 3D person models, the salient characteristic of thisengine is "controllable". That is, we are able to synthesize pedestrians bysetting the visual variables to arbitrary values. Second, on the 3D dataengine, we quantitatively analyze the influence of pedestrian rotation angle onre-ID accuracy. Comprehensively, the person rotation angles are preciselycustomized from 0 to 360, allowing us to investigate its effect on thetraining, query, and gallery sets. Extensive experiment helps us gain deeperunderstanding of the fundamental problems in person re-ID. Our research alsoprovides beneficial insights for dataset building and future practical usage,e.g., a person of a side view makes a better query.

 

Introduction (beta)

None

 

Conclusion (beta)

None