Person Re-Identification via Discriminative Accumulation of Local Features

Tetsu Matsukawa, Takahiro Okabe, Yoichi Sato

Abstract: Metric learning to learn a good distance metric fordistinguishing different people while being insensitive to intra-person variations is widely applied to person re-identification.In previous works, local histograms are densely sampled toextract spatially localized information of each person image.The extracted local histograms are then concatenated into onevector that is used as an input of metric learning. However, thedimensionality of such a concatenated vector often becomes largewhile the number of training samples is limited. This leads toan over fitting problem. In this work, we argue that such aproblem of over-fitting comes from that it is each local histogramdimension (e.g. color brightness bin) in the same position istreated separately to examine which part of the image is morediscriminative. To solve this problem, we propose a methodthat analyzes discriminative image positions shared by differentlocal histogram dimensions. A common weight map shared bydifferent dimensions and a distance metric which emphasizesdiscriminative dimensions in the local histogram are jointlylearned with a unified discriminative criterion. Our experimentsusing four different public datasets confirmed the effectivenessof the proposed method