Interesting papers (semi-supervised training methods)

[rpytel1]

Please post here interesting papers, possibly with a brief description of technique and results (improvement).

[rpytel1]

Semi-supervised semantic segmentation needs strong, high-dimensional perturbations

Interesting method improving 8% on 100 samples in CitySpace

[motykatomasz]

A Probabilistic U-Net for Segmentation of Ambiguous Images

They use VAE to model prior distribution and combine it with U-Net. Tested on Cityscape dataset. Pytorch code: https://github.com/stefanknegt/Probabilistic-Unet-Pytorch

[motykatomasz]

Biasing Deep ConvNets for Semantic Segmentation of Medical Images with a Prior-driven Prediction Function

Simple method for semantic segmentation with prior information. The authors first create spatial prior map and then incorporate it into final softmax layer. Highly relevant for medical imaging but maybe we can somehow incorporate the idea into Cityscape data.

[motykatomasz]

Adversarial Learning for Semi-Supervised Semantic Segmentation

GANs in semantic segmentation setting. The authors trained it on 1/8 Cityscape.

[rpytel1]

Universal Semi-Supervised Semantic Segmentation

• Interesting approach introducing refined entropy module.
• Requires unlabeled dataset.
• 12 % improvement on CityScapes compared to baseline.
• Pytorch code: https://github.com/tarun005/USSS_ICCV19

Semi-Supervised Semantic Segmentation with Cross-Consistency Training

• Interesting perturbation techniques - feature level, prediction-level etc.
• 50 epochs, lr=0.01, poly-learning rate policy,
• Pytorch code: Not available yet :(

Data Efficient Unsupervised Domain Adaptation for Cross-Modality Image Segmentation

• Combines a novel VAE-based feature prior matching, which is data-efficient, and domain adversarial training to learn a shared domain-invariant latent space which is exploited during segmentation.
• Adversarial training 👎
• Pytorch code: didn't share...