NOTE: In this repository, we only implement the VGG16 version.
Requirements
TensorFlow r0.10 (r0.11 should be fine, not sure if this can work for later versions)
OpenCV 2 and its Python bindings
Ipdb: IPython environment debugger
(Optional) pathos. Check the other branch for further details
Prepare dataset
In this implementation, we use the VOC2011 dataset. Please do as follows to set up the dataset:
mkdir data to set up the directory of dataset
Download the train/val dataset and Development kit tar files, put them under the data folder. Unzip Development kit tar file, then unzip train/val tar file and rename the folder as VOC2011.
It should have this basic structure (under data directory)
$ VOCdevkit/ # development kit$ VOCdevkit/VOCcode # VOC utility code$ VOCdevkit/VOC2011 # image sets, annotations, etc.# ... and several other directories ...
You may also download the test set if you want to evaluate your prediction results on this dataset.
Pre-trained model
mkdir model
We use a ImageNet pre-trained model to initialize the network, please download the npy file here and put it under the model folder.
How to train
Since input images have different sizes, in order to make them as minibatch input, we used two different strategies: 1) padding to a large size; or 2) resize to a small size (256, 256)
cd src
python train.py # paddingpython train_small.py # resize
You can choose either one to run. And you can also change the config dictionary to use custom settings.
Demo
cd src
python demo.py
You can change the config dictionary to use custom settings.
Generate Predictions
First you should run the following code:
cd src
python test.py
You might want to change the used model. Check the code for futher details.
After that, you should find the following structure in result folder:
$ FCN8_adam_iter_10000/ # folder name depends on the model you used$ FCN8_adam_iter_10000/gray/ # gray-scale segmentation result$ FCN8_adam_iter_10000/rgb/ # rgb segmentation result# ... and maybe several other directories ...
Then you can use the VOC2011 provided eval code to do evaluation (see the next section for details).
If you want to evaluate your model on the test split, you may submit your prediction results to their server
Evaluation
cd misc
Run save_colorful_grayscale(in_directory,out_directory) (our generated results is grayscale png format, but the eval code uses indexed png format)
Run report.m
Note:
make sure VOCinit.m is at /tf_fcn-master/data/VOCdevkit/VOCcode/
make sure the segmentation result is stored in /tf_fcn-master/data/VOCdevkit/results/VOC2011/Segmentation/%s_val_cls/ while the folder is named as %s_val_cls
make sure the second input of Evaluation(VOCopts, ~) is the string %s above.
Note: When you train the shortcut version model (FCN16 and FCN8), you will need FCN32 model npy file as initialization, instead of the ImageNet pre-trained model npy file.