In this paper we present a simultaneous segmentation algorithm for multiple highly-occluded objects, which combines high-level knowledge and low-level information in a unified framework. The high-level knowledge provi...In this paper we present a simultaneous segmentation algorithm for multiple highly-occluded objects, which combines high-level knowledge and low-level information in a unified framework. The high-level knowledge provides sophis- ticated shape priors with the consideration of blocking relationship between nearby objects. Different from conventional layered model which attempts to solve the full ordering problem, we decompose the problem into a series of pairwise ones and this makes our algorithm scalable to a large number of objects. Objects are segmented in pixel level with higher-order soft constraints from superpixels, by a dual-level conditional random field. The model is optimized alternately by object layout and pixel-wise segmentation. V^e evaluate our system on different objects, i.e., clothing and pedestrian, and show impressive segmentation results and significant improvement over state-of-the-art segmentation algorithms.展开更多
基金supported in part by the National Natural Science Foundation of China under Grant No.61075026the National Basic Research 973 Program of China under Grant No.2011CB302203.
文摘In this paper we present a simultaneous segmentation algorithm for multiple highly-occluded objects, which combines high-level knowledge and low-level information in a unified framework. The high-level knowledge provides sophis- ticated shape priors with the consideration of blocking relationship between nearby objects. Different from conventional layered model which attempts to solve the full ordering problem, we decompose the problem into a series of pairwise ones and this makes our algorithm scalable to a large number of objects. Objects are segmented in pixel level with higher-order soft constraints from superpixels, by a dual-level conditional random field. The model is optimized alternately by object layout and pixel-wise segmentation. V^e evaluate our system on different objects, i.e., clothing and pedestrian, and show impressive segmentation results and significant improvement over state-of-the-art segmentation algorithms.
