While recent advances in deep learning pushed the state-of-the-art in object detection and semantic segmentation, it often comes at the cost of a considerable annotation effort. Thus, weakly supervised learning became of increasing interest. In this paper a novel approach to the challenging task of weakly supervised segmentation and object localization will be presented. The problem is tackled from a mixed perspective utilizing a discriminative and a generative approach. A kernel density estimation is initialized based on the output of a fully convolutional network that is trained in a weakly supervised manner. The proposed approach is evaluated on the VOC2012 dataset for the task of semantic segmentation and object localization. The results show a great improvement is achieved by combining the output of the discriminatively learned network with a subsequent generative segmentation model.