<i>n</i> -CPS: Generalising Cross Pseudo Supervision to <i>n</i> Networks

Cross pseudo supervision (CPS) is an effective method for semi-supervised semantic segmentation, which trains two neural networks with a custom cross supervision scheme. As we observe that it uses only one of those networks in the inference phase, we suggest utilizing both networks using voting and generalising it to more than two networks. As a result, we present <inline-formula> <tex-math notation="LaTeX">$n$ </tex-math></inline-formula>-CPS, a generalization of CPS that uses <inline-formula> <tex-math notation="LaTeX">$n$ </tex-math></inline-formula> simultaneously trained subnetworks that learn from each other through one-hot encoding perturbation and consistency regularization, which together with ensembling of the trained subnetworks significantly improves the performance over the prior method. Paired with CutMix, <inline-formula> <tex-math notation="LaTeX">$n$ </tex-math></inline-formula>-CPS (<inline-formula> <tex-math notation="LaTeX">$n=3$ </tex-math></inline-formula>) consistently outperformed CPS for Pascal VOC 2012 and Cityscapes for 1/16, 1/8, 1/4, and 1/2 supervised regimes. Our method has also served as the foundation for the more recent Diverse Co-Training, which together with <inline-formula> <tex-math notation="LaTeX">$n$ </tex-math></inline-formula>-CPS holds the best results for numerous ResNet-50 benchmarks on these datasets.