semantic_segmentation_implementation

PyTorch implementation of semantic segmentation models

Models

• FCN (FCN-8s)
• DeconvNet
• PSPNet (ResNet18 bottleneck)

Dependencies

• Pytorch
• torchvision
• numpy
• PIL
• tqdm
• click

Dataset Settings

I used VOC2012 dataset with additonal labeled data.

.
└── data
    └── VOCdevkit
         └── VOC2012
            ├── JPEGImages
            |    ├── 2007_000027.jpg
            |    └── ...
            └── SegmentationClassAug
                 ├── 2007_000027.png
                 └── ...

Train & Test

python train.py --model_name=[fcn, deconvnet, pspnet]
python test.py --model_name=[fcn, deconvnet, pspnet] --image_path=(PATH)

Result

• PSPNet evaluation with PSPNet trained 36 epochs (loss: 0.47 ~ 0.50)

Input image	True label	Estimated

Reference

• https://github.com/HyeonwooNoh/DeconvNet
• https://github.com/Lextal/pspnet-pytorch