Multiple Kernel Boosting Based Two-level RGBD Image Co-Segmentation

摘要

RGBD image co-segmentation aims at automatically discovering and segmenting common objects from a set of relevant RGBD images. In this paper, we propose a novel RGBD image co-segmentation method which takes advantage of sample selection as well as multiple kernel boosting (MKB) classifiers. First, the initial co-segmentation results are generated based on RGBD co-saliency maps. Then the sample selection is performed on the basis of pre-segmented regions according to their mutual similarity. The most similar regions with high co-saliency values are further picked out from different images in the image set. Subsequently, on the region level, all positive and negative samples selected above are used to train the MKB classifier for the whole image set. Meanwhile, on the pixel level, the samples selected in each image are exploited to learn a pixel-level MKB classifier. Finally, the co-segmentation results are generated by fusing the results of region-level co-segmentation results and pixel-level segmentation results as well as initial co-segmentation results. Experimental results on a public RGBD image co-segmentation dataset demonstrate that the proposed co-segmentation method outperforms the state-of-the-art co-segmentation methods.