Learning to segment with image-level annotations

摘要: Recently, deep convolutional neural networks (DCNNs) have significantly promoted the development of semantic image segmentation. However, previous works on learning segmentation network often rely a large number ground-truths with pixel-level annotations, which usually require considerable human effort. In this paper, we explore more challenging problem by to segment under image-level annotations. Specifically, our framework consists two components. First, reliable hypotheses based localization maps are generated incorporating hypotheses-aware classification and cross-image contextual refinement. Second, can be trained in supervised manner these maps. We training strategies for achieving good performance. For first strategy, novel multi-label cross-entropy loss is proposed train directly using multiple all classes, where each pixel contributes class different weights. second rough mask inferred from maps, then optimized single-label produced masks. evaluate methods PASCAL VOC 2012 benchmark. Extensive experimental results demonstrate effectiveness compared state-of-the-arts. HighlightsLocalization map generation hypothesis-based classification.A maps.An effective method predict given image.Our achieve new state-of-the-art