Evaluation framework for testing segmentation networks in PyTorch. What segmentation network to choose for next Kaggle competition? This benchmark knows the answer!

This repository is not maintained. Please refer to https://github.com/BloodAxe/pytorch-toolbelt instead.

It tries to show pros & cons of many existing segmentation networks implemented in Keras and PyTorch for different applications (biomed, sattelite, autonomous driving, etc). Briefly, it does the following:

for model in [Unet, Tiramisu, DenseNet, ...]:
    for dataset in [COCO, LUNA, STARE, ...]:
        for optimizer in [SGD, Adam]:
            history = train(model, dataset, optimizer)
            results.append(history)
summarize(results)

• Write Keras train pipeline
• Write Pytorch train pipeline

• Add ZF_UNET model (https://github.com/ZFTurbo/ZF_UNET_224_Pretrained_Model)
• Add LinkNet model
• Add Tiramisu model (https://github.com/0bserver07/One-Hundred-Layers-Tiramisu)
• Add SegCaps model
• Add VGG11,VGG16,AlbuNet models (https://github.com/ternaus/TernausNet)
• Add FCDenseNet model (https://github.com/bfortuner/pytorch_tiramisu)

• Add DSB2018 (stage1) dataset
• Add COCO dataset
• Add STARE dataset
• Add LUNA16 dataset
• Add Inria dataset
• Add Cityscapes dataset
• Add PASCAL VOC2012 dataset

• Add fancy plots

• https://github.com/ZFTurbo/ZF_UNET_224_Pretrained_Model
• https://github.com/ternaus/TernausNet
• https://github.com/0bserver07/One-Hundred-Layers-Tiramisu
• https://github.com/bfortuner/pytorch_tiramisu
• https://raw.githubusercontent.com/ZijunDeng/pytorch-semantic-segmentation
• https://github.com/mapillary/inplace_abn