Semantic graph construction for weakly-supervised image parsing

摘要: We investigate weakly-supervised image parsing, i.e., assigning class labels to regions by using imagelevel only. Existing studies pay main attention the formulation of learning problem, how propagate from images given an affinity graph regions. Notably, however, regions, which is generally constructed in relatively simpler settings existing methods, crucial importance parsing performance due fact that problem cannot be solved within a single image, and enables label propagation among multiple images. In order embed more semantics into graph, we propose novel criteria exploiting weak supervision information carefully, develop two graphs: L1 semantic k-NN graph. Experimental results demonstrate proposed graphs not only capture relevance, but also perform significantly better than conventional parsing.