Building a graph model use the whole training set and solved by graph cut based algorithm is a common method in weak supervision semantic segmentation task, such as Multi-Image Model (MIM). It has two disadvantages: one is the parameter number of model increased rapidly with the scale growth of training set, which limited applied to large-scale data. Another is lack of use structure information in image internal. To solve above problems, we proposed a Constrained Multi-Image Model (CMIM) that training model with a part of the training data which acquired by our entropy based algorithm. It's made up of some components and each is a smaller graph. So, The CMIM can parallel or serial training and weaken the memory limit. To utilize the context information, we bring the saliency of image to unary potential in energy function. At first, we segment images to superpixels and extract the semantic texton forest (STF) feature. Then construct a conditional random fields (CRF) in the superpixel set from selected images. The data potential learned from STF featrue and saliency of superpixels. Finally, the labeling of superpixels converted to CRF optimization problem which can efficiency solved by alpha expansion algorithm. Experiments on the MSRC21 dataset show that the CMIM algorithm achieves accuracy comparable with some previous influential weakly-supervised segmentation algorithms.
Weakly-supervised semantic segmentation is a challenge in the field of computer vision. Most previous works utilize the labels of the whole training set and thereby need the construction of a relationship graph about image labels, thus result in expensive computation. In this study, we tackle this problem from a different perspective. We proposed a novel semantic segmentation algorithm based on background priors, which avoids the construction of a huge graph in whole training dataset. Specifically, a random forest classifier is obtained using weakly supervised training data .Then semantic texton forest (STF) feature is extracted from image superpixels. Finally, a CRF based optimization algorithm is proposed. The unary potential of CRF derived from the outputting probability of random forest classifier and the robust saliency map as background prior. Experiments on the MSRC21 dataset show that the new algorithm outperforms some previous influential weakly-supervised segmentation algorithms. Furthermore, the use of efficient decision forests classifier and parallel computing of saliency map significantly accelerates the implementation.