3M: An Affinity Processing Method for Weakly Supervised Semantic Segmentation

In the face of the high cost of annotation brought by fully supervised semantic segmentation, a way to simplify the annotation has been proposed—weakly supervised semantic segmentation. Traditional weakly supervised semantic segmentation mostly adopts multi-stage training approach, resulting in more complex and time-consuming training of a complete segmentation network. In this paper, a single-stage end-to-end weakly supervised semantic segmentation is proposed to reduce the complexity of segmentation network training. The traditional weakly supervised semantic segmentation network based on convolutional neural network can not overcome the loss of global contextual information brought by the convolutional kernel. In this paper, we use the transformer-based backbone network to effectively retain the contextual information. For the inadequacy of the existing network to deal with the semantic affinity information module, this paper puts forward a statistically empirical affinity map processing method named 3M to improve the segmentation accuracy. This paper conducts extensive experiments and shows better results on Pascal VOC12 dataset.