Superpixel-based background removal for accuracy salience person re-identification

This paper presents superpixel-based background removal methods to increase accuracy of global salience person re-identication method. The current algorithm has two problems which limit its accuracy including (1) wrong matching when images of different people have the same background and/or (2) salience on the background of different images of different people is similar. Theoretical maximum accuracy of the global salience method when applying background removal was proved in our previous work by manually remove background close to a human. To achieve an automatic background removal with an accuracy close to the theory, we propose two approaches namely (i) superpixel-GBVS method that combines superpixel and local salience information and (ii) superpixel-pose method that combines superpixel and pose estimation information. The two technologies decide what superpixels belong to human and the others belong to background. It results in a background removal close to edges of the human. Preliminary results show that our results outperform those of the original method without applying any background removal methods and results of the second approach are close to the theory.