This repo is pet project in computer vision and semantic segmentation. The main aim is design or adaptate CNN to segment windows on different buildings and count them. I used CMP Facade Dataset for this task. Dataset contains 606 images with corresponding masks.
cd your_path
git clone https://github.com/DmitriyKras/facade-segmentation.git
cd facade-segmentation
mkdir weights
mkdir weights/checkpoint
mkdir data
Config files prepare_config.json, train_config.json and evaluate_config.json contain all important params such as pathes, training settings etc for experiment process. Modify them according to provide information you need. Note: all pathes are relative to project's root folder. utils and models folders contain py files for training, data preparation and model building. src folder contains all files for experiment such as src/prepare.py, src/train.py and src/evaluate.py. pairs folder used to store processed train, val and test pairs.
By default weights saved after training and used for evaluation stored in weights, weights saved during training stored in weights/checkpoint.
Resulting JSON files with training and evaluation results stored in results/training and results/evaluation respectively, figures with PR and ROC curves stored as PNG in results/evaluation. Images and masks stored in data/images and data/masks.
CMP Facade dataset has 12 classes with different parts of facade. We are interested in only windows, so utils/process_masks.py used to prepare masks for our task. Now all masks contain 255 at windows and 0 at other locations. Processed dataset is available at my Kaggle profile: ... . To generate train, val and test pairs make changes at prepare_config.json to provide relative pathes to images and masks folders and destination folder where generated pairs will be stored as JSON files. Then run src/prepare.py
Make changes to train_config.json to provide pathes to train and val pairs and other settings such as batch size, epochs etc.
Then run src/train.py to start training process, JSON file with training results will be generated and weights will be saved according to config file.
Make changes to eval_config.json to provide required evaluation information and run src/evaluate.py.
JSON file with evaluation results and figures with ROC and PR curves will be generated according to config file.
Here are examples of figures with ROC and PR curves generated by evaluate.py with model trained 20 epochs.
