Weakly supervised image parsing by discriminatively semantic graph propagation

In this paper, we concentrate on a challenging problem - image parsing trained on images with weakly supervised information, i.e., image-level labels. Image-level labels are ambiguous and difficult for training. Typically, an affinity graph of superpixels is constructed to provide additional information about labels of the target superpixel. However, existing work constructs affinity graph in a naive manner, L1 reconstruction and k-NN are most used where label co-occurrence is a common phenomenon and degenerates the assignment performance. To overcome above problem, we proposed the use of discriminatively semantic ability between neighbor superpixels and the target superpixel in affinity graph construction. With simpler experiment setup and lower time complexity, our method achieves average per-class accuracy comparable to state-of-the-art performances in weakly-supervised image parsing task on datasets MSRC-21 and PASCAL VOC 2007.