/pytorch-semantic-segmentation

PyTorch implementation of FCN-based models for semantic segmentation

Primary LanguagePython

PyTorch implementation of FCN-based models for semantic segmentation

This repository contains some FCN-based models and the pipeline of training and evaluating models.

Models

  1. Vanilla FCN: FCN32, FCN16, FCN8, in the versions of VGG19, ResNet152 and DenseNet201 respectively.
  2. U-Net
  3. SegNet

Usage

  1. Modify the configuration.py according to the hint in it.
  2. Run split_train_val.py.
  3. Set your model and training parameters in train.py and then run.

Observation

  1. Weight initialization is very important, without which the training even cannot converge no matter what learning rate is used.

Reference

  1. I have referred to some nice repositories: [1], [2]

TODO

  1. PSPNet
  2. DeepLab
  3. CRFAsRNN
  4. More dataset
  5. And so on