Background removal for improving saliency-based person re-identification
Thuy Binh Nguyen, Pham Van Phu, Thi‐Lan Le, Cuong Vo Le · 2016
This paper presents background removal methods to increase the accuracy of saliency-based person re-identification. After evaluating the current global salience algorithm, we found that wrong matching appears when (1) images of different people have a similar or the same background and/or (2) salience on the backgrounds of various images are similar. To prove the maximum theoretical accuracy of the global saliency method when using background removal, we use a manual method with support of an interactive segmentation tool. Another method is to use an ellipse to localize a human body region. This method is a preliminary step for confirming a possibility of applying an automatic technique. The human body region is automatically determined by utilizing a local salience method named GBVS with an adaptive threshold for every image in VIPeR dataset. Preliminary results show that when applying the three background removal methods, the accuracy at rank 1 of CMC curve increases from 20.00% to 27.18%, 24.81% and 23.80% for interactive segmentation, elliptical window and adaptive local salience methods, respectively.