This is my undergraduate dissertation project. The goal of this thesis is to examine and compare the results from two variations of CNN Encode-Decode arhitectures using Self-Normalization technique along with CRF-RNN post processing unit. Due to visualize the results of the model properly a Visualizer based on CityscapesScripts has been implemented to visualize the results.
- python 2.7
- keras 2.1
- tensorflow 1.4
- scikit-learn 0.19
- openCV 2.4
- numpy 1.13
- scipy 0.13
- pyQt4 for the Visualizer
Run pip install -r requirements.txt to intall the dependencies
train.py [-h] [-n NETWORK] [-trp TRAINPATH] [-vdp VALIDATIONPATH] [-tsp TESTPATH] [-bs BATCHSIZE] [-crf] [-w [WEIGHTS]] [-m [MODEL]] [-e EPOCHS]
Run make inside lib/crfasrnn_keras/src/cpp to build highdimfilter module.
Create the npy data files for the data generator using denseExtraction.py.
Check the examples below to train your model.
python train.py -n bdcnn -trp trainpath -vdp validationpath -tsp testpath -bs 4 -crf -e 20
python train.py -trp trainpath -vdp validationpath -tsp testpath -bs 4 -w weightspath -m modelpath -e 20
- I want to thank Sadeep Jayasumana for his excellent work with CRF-RNN post-processing unit implementation in keras.