Dense FixMatch: a simple semi-supervised learning method for pixel-wise prediction tasks

We propose Dense FixMatch, a simple method for online semi-supervised learning of dense and structured prediction tasks combining pseudo-labeling and consistency regularization via strong data augmentation. We enable the application of FixMatch in semi-supervised learning problems beyond image classification by adding a matching operation on the pseudo-labels. This allows us to still use the full strength of data augmentation pipelines, including geometric transformations. Figure 1: Dense FixMatch (blue) on unlabeled data We evaluate it on semi-supervised semantic segmentation improves the performance of semi-supervised semantic on Cityscapes and Pascal VOC with different segmentation on Cityscapes val set using percentages of labeled data and ablate design DeepLabv3+ with ResNet-101 backbone over supervised choices and hyper-parameters. Dense FixMatch baselines (red) across different amounts of significantly improves results compared to supervised labeled samples.