Interesting papers (models & losses)

[rpytel1]

Interesting resources with an intro to image segmentation field:

• Image segmentation in 2020: Architectures, Losses, Datasets, and Frameworks
• A 2019 Guide to Semantic Segmentation

Models

UNet (MICCAI, 2015)

• already implemented in fast.ai
• one of the most popular models in Image Segmentation
• kind of old architecture

Tiramisu (2017)

• code available in fast.ai course materials
• 88 % accuracy on Camvid

FastFCN(2019)

• newer architecture
• code available (Pytorch code: https://github.com/wuhuikai/FastFCN)
• fast.ai code: TBA
• fully connected network with joint pyramid upsampling

Gated-SCNN(2019)

• takes two types of input (casual image and shapes)
• one of the newest architectures
• code (Pytorch implementation: https://github.com/nv-tlabs/GSCNN)
• fast..ai code: TBA

Losses

Focal loss

• improvement of cross-entropy loss
• Pytorch implementation (https://www.kaggle.com/c/tgs-salt-identification-challenge/discussion/65938)

Dice loss

• most commonly used loss in Image Segmentation
• Pytorch code (https://github.com/kevinzakka/pytorch-goodies/blob/master/losses.py)

Boundary loss

• advised for unbalanced datasets
• Pytorch implementation of this and other losses (https://github.com/JunMa11/SegLoss/tree/master/losses_pytorch)

Lovász-Softmax loss

• This loss performs direct optimization of the mean intersection-over-union loss in neural networks based on the convex Lovasz extension of sub-modular losses.
• Pytorch implementation (https://github.com/bermanmaxim/LovaszSoftmax/blob/master/pytorch/lovasz_losses.py)