pytorch-fcn

Fully Convolutional Networks implemented with PyTorch.

Requirements

• pytorch >= 0.2.0
• torchvision >= 0.1.8
• fcn >= 6.1.5
• Pillow
• scipy
• tqdm

Installation

git clone https://github.com/wkentaro/pytorch-fcn.git
cd pytorch-fcn

conda install pytorch cuda80 torchvision -c soumith
pip install .

Training

See VOC example.

Accuracy

At pytorch-fcn==1.7.0.

Model	Implementation	epoch	iteration	Accuracy	Accuracy Class	Mean IU	FWAV Accuracy
FCN32s	Original	-	-	90.49	76.48	63.63	83.47
FCN32s	Ours	10	92000	90.63	72.36	63.13	83.36
FCN16s	Original	-	-	91.00	78.07	65.01	84.27
FCN16s	Ours	5	44000	91.03	77.38	64.80	84.23
FCN8s	Original	-	-	91.23	77.61	65.51	84.55
FCN8s	Ours	3	28000	91.24	77.12	65.39	84.55
FCN8sAtOnce	Original	-	-	91.13	78.50	65.40	84.44
FCN8sAtOnce	Ours	6	56000	91.12	76.42	65.10	84.36

Visualization of validation result of FCN8s.