WHAT IS FOREGROUND: FROM THE VIEW OF PHOTOGRAPHERS
Hong Lu, Qian Diao, Renzhong Wei, Cheng Jin, Xiangyang Xue
Shanghai Key Lab of Intel. Info. Processing
School of Computer Science, Fudan University
-ABSTRACT-
In this paper we focus on foreground extraction in digital images. We redefine foreground as an object close to the camera in real distance, and at the same time clear in the image. Based on this observation, we propose a depth of field measure which represents the distance of object from camera, and cooperate with saliency based approach. Specifically, for a given image, we firstly extract a Saliency Map (SM) and a Defocus Map(DM), with their value corresponding to saliency value and depth of field value. Then we segment images and extract segment feature based on these two extracted maps. A classifier is trained to separate foreground from background. On test images, segments are considered to be foreground if they have high enough confidence on the trained classification model. The proposed method is tested on MSRA Salient Object Detection image set and Flickr image set. Experimental result demonstrates that our method can obtain better result on foreground extraction over related methods.
-Images Set-
MSRA-SOD –download here
Flickr – download here
-Code-
Code – download here
-Results-
