Tensorflow-SegNet

Implement slightly different (see below for detail) SegNet in tensorflow, successfully trained segnet-basic in CamVid dataset.

Due to indice unravel still unavailable in tensorflow, the original upsampling method is temporarily replaced simply by deconv( or conv-transpose) layer (without pooling indices). You can follow the issue here: https://github.com/tensorflow/tensorflow/issues/2169 (The current workaround for unpooling layer is a bit slow because it lacks of GPU support.)

for model detail, please go to https://github.com/alexgkendall/caffe-segnet

Requirement

tensorflow 1.0 Pillow (optional, for write label image) scikit-image

Update

Update to tf 1.0

Finally get some time to refactor a bit, removing some un-used function and remove the hard-coded file path Now the model should be easy to config. The parameters can be found in main.py.

I planned to add more feature such as dilation, multi-resolution, sequential learning..etc. Making it more like a "basic" segmentation toolbox and support more dataset as well. Therefore the model and documentation will be changed accordingly in the future.

More utility function will be added and some messed coding style will be fixed. Any feature request is also welcomed.

Usage

see also example.sh training:

python main.py --log_dir=path_to_your_log --image_dir=path_to_CamVid_train.txt --val_dir=path_to_CamVid_val.txt --batch_size=5

finetune:

python main.py --finetune=path_to_saved_ckpt --log_dir=path_to_your_log --image_dir=path_to_CamVid_train.txt --val_dir=path_to_CamVid_val.txt --batch_size=5

testing:

python main.py --testing=path_to_saved_ckpt --log_dir=path_to_your_log --test_dir=path_to_CamVid_train.txt --batch_size=5 --save_image=True

You can set default path and parameters in main.py line 6~18. note: in --testing you can specify whether to save predicted images, currently only save one image for manually checking, will be configured to be more flexible.

Dataset

This Implement default to use CamVid dataset as described in the original SegNet paper, The dataset can be download from author's github https://github.com/alexgkendall/SegNet-Tutorial in the CamVid folder

example format:

"path_to_image1" "path_to_corresponded_label_image1",

"path_to_image2" "path_to_corresponded_label_image2",

"path_to_image3" "path_to_corresponded_label_image3",

.......