PixelLink-with-pytorch

PixelLink-with-pytorch

Prerequisite

• python-3.6
• pytorch-0.4.0
• torchvision-0.2.1
• opencv-3.4.0.14
• numpy-1.14.3
• Pillow-5.5.0

They could all be installed through pip except pytorch and torchvision. As for pytorch and torchvision, they both depends on your CUDA version, you would prefer to reading pytorch's official site

Structure

All main source code is in the root directory of the project.

• ${project_root}/unittest contains code you could run indenpendently, which identifies some modules's function.
• datasets.py contains code which generates datasets and preprocess code
• net.py contains the neural network structure
• criterion.py contains code which calculates the loss
• postprocess.py contains code for data postprocessing, which transform pixel and link mask to bounding boxes
• config.py contains almost all changeable parameters.
• other *.py are useless, they exists only for re-constructing the project later on.

Train

Before starting

You could modify training parameters in ${project_root}/config.py You need to download the dataset here and unzip it as ${project_root}/train_images/images and ${project_root}/train_images/ground_truth python main.py --train 1

Retrain

Be sure there is a pretrained model in the ${project_root}/models directory python main.py --train 1 --retrain 1

Test

Be sure there is a pretrained model in the ${project_root}/models directory python main.py

Noted

There are still some bugs in source code. The result is not satisfactory. Still under developing...