Semisupervised Semantic Segmentation with Mutual Correction Learning

Figure 1

Overview of mutual correction learning. Two images and are sampled from the unlabeled dataset. The CutMix images are generated by two source images, and they are all inputted into each segmentation network. and are mixed as pseudo segmentation maps to supervise the other segmentation network. : CutMix, MFFA: multiscale feature fusion attention mechanism module, : mutual correction loss, : cross pseudo supervision loss, : segmentation confidence map, : predicted one-hot label map, and SE module: squeeze-and-excitation module.