Abstract

Recognizing objects from disjoint camera views, known as person re-identification, is an important and challenging problem in the field of computer vision. Recent progress in person re-identification is due to new visual features and models that deal with cross-view differences. Existing appearance models focus on visual features in the normal sense, e.g., color histogram, Scale-invariant Feature Transform (SIFT) and Histogram of Oriented Gradients (HOG). In this paper, we propose a new appearance based method using the generative information of local image features and their encoding. In this paradigm, local image features which capture the color and structural cues of the human images are first extracted. A Gaussian Mixture Model (GMM) is then learned to approximate the generation process of these features. It provides a relatively comprehensive statistical representation. Finally, discriminative feature maps are obtained by calculating Free Energy Score Space (FESS) for GMM. The obtained feature maps are concatenated and encoded into a fixed-length feature vector for person re-identification. Our approach demonstrates promising performance on challenging datasets. It is also very practical: it has low computational cost both at training and testing. A GMM trained on images with different imaging conditions can be applied to other images without any significant loss in performance.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call