We consider the task of learning a classifier for semantic segmentation using weak supervision, in this case, image labels specifying the objects within the image. Our method uses deep convolutional neural networks (CNNs) and adopts an Expectation-Maximization (EM) based approach maintaining the uncertainty on pixel labels. We focus on the following three crucial aspects of the EM based approach: (i) initialization; (ii) latent posterior estimation (E step) and (iii) the parameter update (M step). We show that {\em saliency} and {\em attention} maps provide good cues to learn an initialization model and allows us to skip the bad local minimum to which EM methods are otherwise traditionally prone. In order to update the parameters, we propose minimizing the combination of the standard \textit{softmax} loss and the KL divergence between the true latent posterior and the likelihood given by the CNN. We argue that this combination is more robust to wrong predictions made by the expectation step of the EM method. We support this argument with empirical and visual results. We additionally incorporate an approximate intersection-over-union (IoU) term into the loss function for better parameter estimation. Extensive experiments and discussions show that: (i) our method is very simple and intuitive; (ii) requires only image-level labels; and (iii) consistently outperforms other weakly supervised state-of-the-art methods with a very high margin on the PASCAL VOC 2012 dataset.

本研究提出一种使用CNNs结合基于期望最大化算法的弱监督方法来学习语义分割分类器的方法，并通过在简单图像上学习初始化和使用显著性和注意力地图来实现更好的结果，同时，研究也表明该方法无需使用精确像素级别标签，能在PASCAL VOC 2012数据集上保持高性能。

基于自下而上和自上而下的线索进行弱监督语义分割