This work was finished for a practice purpose. https://github.com/wudonghao/Hand-Segmentation.git

Customized implementation of the U-Net in Pytorch. This was used with only one output class. The model used for the last submission is stored in the MODEL.pth file. It was pretrained with a modle available on the Kaggle website, yet which is originally for car segmentation. The model is scored a dice coefficient of 0.90 on a part of images from GTEA_gaze dataset.

pytorch version: 0.3.1

To roll back your pytorch, you can run only with

"pip uninstall torch"

"pip install torch==0.3.1"

Try " python test.py -h " to see more options.

Simply test one image and rename the output by running

" python test.py -i test1.jpg -o ouput1.jpg "

Test multiple images by running

" python test.py -i test1.jpg test2.jpg --model MODEL.pth "

See the result in a plot by running

" python test.py -i test1.jpg --viz "

Try " python train.py -h " to see more options. (Default value of val_train percentage equals to 0.05)

Simply train by running

"python train.py"

Train with GPU model

"python train.py -g"

Unzip "result.tar.gz" to get all the saved test results.

They can be evaluated by using dice coefficient by running "python eval_all.py -b 2 -l 0.02 -c MODEL.pth -g"

Run "bash download_data.sh" to get the dataset. (around 700M)

Run "bash download_model.sh" to get the model. (around 50M)

1.Notice that dirs in this work are absolute path. Please change them for your usage.

2.In order to process the image, it is split into two squares (a left on and a right one), and each square is passed into the net. The two square masks are then merged again to produce the final image. As a consequence, the height of the image must be strictly superior than half the width. Make sure the width is even too.

Images for training in this work are only avaliable with the size of 1918 * 1280. Please resize your own data in advance or optimize this project for your convenience.

3.Batch_size over 4 causes overflow on my machine probably for the above reason.

4.The ground truth images as input have been binarized with 0 and 1 during preprocessing.

5.Some codes for image processing are provided in the file named "Image_processing_tools" while without any guarantees for using directly.

This package depends on pydensecrf, available via pip install.