A Simple Deep Feature Representation for Person Re-identification

Abstract

Person re-identification (Re-ID) aims to match persons across non-overlapping camera views at different time. Typical person Re-ID models include two critical components: feature representation and metric learning. Due to the large variations in a persons appearance by different poses, viewpoints, illumination and occlusions, metric learning is always a necessary part in person Re-ID. In this paper, we propose a Deep person Feature Representation (DFR) learning frame-work based on a classification-oriented convolution neural network, and the DFR is directly used to calculate cosine distance for the similarity measure while with-out explicit metric learning. In the framework, Batch Normalization (BN) is applied before the ReLU layer to accelerate the convergence process, and with dropout strategy the DFR is only 64-dimension which makes the feature representation more effective and less noisy. Experiments demonstrate that our approach achieves the state-of-the-art results on most of the challenging datasets, especially on dataset of the largest scale CUHK03.