A dynamic portrait segmentation by merging colors and depth information

Abstract

In this paper, we proposed a new method to improve the dynamic portrait segmentation in Kinect which always causes the problem of incomplete image segmentation of portrait due to the loss of the depth. This problem can be solved by using the color information to reinforce the areas where the depth is uncertain. We can segment the portrait's foreground more completely using the proposed method. First, the depth information can be divided into foreground, background, uncertain areas to produce a judging area for the foreground's uncertain areas. Secondly, the volunteer image will be segmented by Sobel edge detection, watershed and other steps in color information then be treated as the characteristic value of color area to calculate the mean value and standard deviation respectively. Finally, we chose the best image from these processing by comparing the color feature of the foreground edge and the judging area. The results show that we can completely segment out the portrait image as well as reduce its error rate significantly.

Joint segmentation of color and depth data based on splitting and merging driven by surface fitting

“…Some approaches deal with the close but less general problem of separating the foreground from the background [25,26,27,28,29]. In [29] two likelihood functions, based on color and depth data, are combined together for this task.…”