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

18 Oct 2022  ·  Miquel Martí i Rabadán, Alessandro Pieropan, Hossein Azizpour, Atsuto Maki ·

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. We evaluate it on semi-supervised semantic segmentation on Cityscapes and Pascal VOC with different percentages of labeled data and ablate design choices and hyper-parameters. Dense FixMatch significantly improves results compared to supervised learning using only labeled data, approaching its performance with 1/4 of the labeled samples.

PDF Abstract

Datasets


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Semi-Supervised Semantic Segmentation Cityscapes 12.5% labeled Dense FixMatch (DeepLabv3+ ResNet-101, uniform sampling, single pass eval) Validation mIoU 73.91% # 14
Semi-Supervised Semantic Segmentation Cityscapes 12.5% labeled Dense FixMatch (DeepLabv3+ ResNet-50, uniform sampling, single pass eval) Validation mIoU 73.39% # 15
Semi-Supervised Semantic Segmentation Cityscapes 6.25% labeled Dense FixMatch (DeepLabv3+ ResNet-101, over-sampling, single pass eval) Validation mIoU 71.1% # 9
Semi-Supervised Semantic Segmentation Cityscapes 6.25% labeled Dense FixMatch (DeepLabv3+ ResNet-50, uniform sampling, single pass eval) Validation mIoU 70.65% # 10
Semi-Supervised Semantic Segmentation Cityscapes 93 labeled Dense FixMatch (DeepLabv3+ ResNet-50, over-sampling, single-pass eval) Validation mIoU 65.81 # 3
Semi-Supervised Semantic Segmentation Cityscapes 93 labeled Dense FixMatch (DeepLabv3+ ResNet-101, over-sampling, single pass eval) Validation mIoU 66.97 # 2
Semi-Supervised Semantic Segmentation Cityscapes with extra (no coarse labels) Dense FixMatch (DeepLabv3+ ResNet-50, over-sampling, single pass eval) Validation mIoU 79.98 # 2
Semi-Supervised Semantic Segmentation Cityscapes with extra (no coarse labels) Dense FixMatch (DeepLabv3+ ResNet-101, over-sampling, single pass eval) Validation mIoU 80.82 # 1
Semi-Supervised Semantic Segmentation Pascal VOC 2012 12.5% labeled Dense FixMatch (DeepLabv3+ ResNet-101, over-sampling, single pass eval) Validation mIoU 65.82% # 29
Semi-Supervised Semantic Segmentation Pascal VOC 2012 12.5% labeled Dense FixMatch (DeepLabv3+ ResNet-50, over-sampling, single pass eval) Validation mIoU 62.49% # 31
Semi-Supervised Semantic Segmentation PASCAL VOC 2012 25% labeled Dense FixMatch (DeepLabv3+ ResNet-101, over-sampling, single pass eval) Validation mIoU 72.04 # 19
Semi-Supervised Semantic Segmentation PASCAL VOC 2012 25% labeled Dense FixMatch (DeepLabv3+ ResNet-50, over-sampling, single pass eval) Validation mIoU 69.02 # 20
Semi-Supervised Semantic Segmentation PASCAL VOC 2012 50% Dense FixMatch (DeepLabv3+ ResNet-101, over-sampling, single pass eval) Validation mIoU 74.73% # 13
Semi-Supervised Semantic Segmentation PASCAL VOC 2012 50% Dense FixMatch (DeepLabv3+ ResNet-50, over-sampling, single pass eval) Validation mIoU 71.69% # 14
Semi-Supervised Semantic Segmentation Pascal VOC 2012 6.25% labeled Dense FixMatch (DeepLabv3+ ResNet-101, over-sampling, single pass eval) Validation mIoU 54.85 # 12
Semi-Supervised Semantic Segmentation Pascal VOC 2012 6.25% labeled Dense FixMatch (DeepLabv3+ ResNet-50, over-sampling, single pass eval) Validation mIoU 52.15 # 13

Methods