Weakly supervised semantic segmentation with a multi-image model

摘要: We propose a novel method for weakly supervised semantic segmentation. Training images are labeled only by the classes they contain, not their location in image. On test instead, predicts class label every pixel. Our main innovation is multi-image model (MIM) - graphical recovering pixel labels of training images. The connects superpixels from all data-driven fashion, based on appearance similarity. For generalizing to new we integrate them into MIM using learned multiple kernel metric, instead learning conventional classifiers recovered labels. also introduce an “objectness” potential, that helps separating objects (e.g. car, dog, human) background grass, sky, road). In experiments MSRC 21 dataset and LabelMe subset [18], our technique outperforms previous methods achieves accuracy comparable with fully methods.