search

usuyama / pytorch-unet

Simple PyTorch implementations of U-Net/FullyConvNet (FCN) for image segmentation
https://colab.research.google.com/github/usuyama/pytorch-unet/blob/master/pytorch_unet_resnet18_colab.ipynb
MIT License
827 stars 233 forks source link

making label #14

Open zengjianyou opened 4 years ago

zengjianyou commented 4 years ago

Your implementation is very good, why don't you use the background as a segmentation category? FCN uses one-hot is often to use the background as a category

usuyama commented 4 years ago

Thanks for your comment. The synthetic data example in this repo uses multi-label segmentation - i.e. objects can be overlapped so one pixel can have multiple positive labels. For multi-class segmentation / one-hot setting, I think it's a good idea to have a background category.