These are the trained models that were used in the papers:
- erfnet_pretrained.net ERFNet full model with encoder trained on ImageNet and decoder trained on Cityscapes train set.
- erfnet_scratch.net ERFNet full model with encoder trained from scratch on Cityscapes train set (1/8 labels) and decoder trained on Cityscapes train set.
- pretrained_encoder_imagenet.net ERFNet encoder trained on ImageNet and used to train '_pretrained' model. Can be used to train ERFNet decoder with the training code by specifying this file in the '--CNNEncoder' option.