semantics_segmentation_of_urban_environments
semantics_segmentation_of_urban_environments copied to clipboard
dimimal
Image segmentation for self-driving car. Based on Cityscapes dataset.
Semantics Segmentation of Urban Environments
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.
Cityscapes Dataset
Dependencies
- 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
Arguments
train.py [-h] [-n NETWORK] [-trp TRAINPATH] [-vdp VALIDATIONPATH] [-tsp TESTPATH] [-bs BATCHSIZE] [-crf] [-w [WEIGHTS]] [-m [MODEL]] [-e EPOCHS]
Results
Installation
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.
Examples
Training
python train.py -n bdcnn -trp trainpath -vdp validationpath -tsp testpath -bs 4 -crf -e 20
Resume Training
python train.py -trp trainpath -vdp validationpath -tsp testpath -bs 4 -w weightspath -m modelpath -e 20
Acknowledgments
- I want to thank Sadeep Jayasumana for his excellent work with CRF-RNN post-processing unit implementation in keras.
dimimal