Weakly-Supervised Semantic Segmentation Network With Iterative dCRF

This Autonomous driving methods driven by big data are becoming more and more perfect, but the cost of existing data labeling is too high, so how to reduce or even not label data has attracted more and more attention. Semantic segmentation networks supervised by image-level annotations are all trained using pseudo-labels. Most methods use image classification networks to generate class activation maps (CAMs) and start with CAMs to diffuse features to other parts of the target to obtain pseudo-labels. However, due to its weak supervision information, it is difficult for the existing methods to obtain better results. Therefore, we propose a weakly-supervised semantic segmentation network with iterative dCRF based on graph convolution. Specifically, we use ResNet to generate CAMs and node features and then use graph convolution for feature propagation and merge the low-level and high-level semantic information of the image. Then execute dCRF in an iterative manner, and finally obtain refined pseudo-labels. On the PASCAL VOC 2012 data set, our model achieves an mIoU of 63.5%, which is 0.3% higher than the graph convolutional network method.

Automated summary

A weakly-supervised semantic segmentation network utilizing graph convolution and iterative dCRF was proposed in this study, achieving high accuracy by combining CAMs and node features generated by ResNet and using graph convolution for feature propagation.