A simple pytorch implement of Fully Convolutional Networks. It use common dependencies and don't need to be built.

Requirements

• python3
• pytorch >= 1.1.0
• numpy, scipy, pillow(pil), scikit-image, opencv, tqdm

pip install -r requirements.txt

Test on Python 3.8 and PyTorch 1.4.

Usage

Training

Run python train.py --model fcn32s --dataset voc --max-iter 20000 --save-dir ./checkpoints/fcn32s. Or sh ./scripts/train_fcn_sbd.sh or sh ./scripts/train_fcn_voc.sh.

Arguments:

• --model: fcn32s or fcn16s or fcn8s.
• --pretrained-model: Path to a pretrained checkpoint. fcn16s(fcn8s) need to initialize from fcn32s(fcn16s) pretrained model. fcn32s need to initialize from VGG pretrained model.
• --batch-size: Batch size of train or val. If set to 1, image and label will be returned as is, otherwise image and label will be resized to 500×500.
• --max-iter: Max training iterations, max epoch will be auto calculate by max-iter / batchsize.
• --lr, --momentum, --weight-decay: Optimizer configs, lr for bias will double.
• --cuda: Whether to use GPU training.
• --dataset: Use VOC 2012 Dataset (1464train/1449val) or SBD Dataset (8498train/736val).
• --dataset-root: Path to dataset.
• --save-dir: Path to save checkpoints, visulized results, and log file.
• --resume: Path to a checkpoint file used to resume training process.

Note:

For VGG pretrained model, torchvision auto downloaded model require input image value between [0, 1] and then normalized with mean=[0.485, 0.456, 0.406],std=[0.229, 0.224, 0.225] in RGB format, as in pytorch/examples; while caffe pretrained model require input image value between [0, 255] and then substracted by [103.939, 116.779, 123.68] in BGR format.

PyTorch version caffe pretrained model can be found at jcjohnson/pytorch-vgg.

Direct download SBD Dataset here.

Demo

Run python demo.py --model fcn32s --checkpoint ./checkpoints/fcn32s/model_best.pth.tar --input-dir ./demo

Results

[mIoU]	FCN-32s	FCN-16s	FCN-8s
VOC	57.86	59.74	60.14
SBD	62.64	64.48	64.87
Original CVPR	59.4	62.4	62.7
Original PAMI	63.6	65.0	65.5

See ./checkpoints dir for more details.

References

• shelhamer/fcn.berkeleyvision.org
• wkentaro/pytorch-fcn
• facebookresearch/maskrcnn-benchmark