This program was my experiment in Deep Learning technology.
My task was to solve the detection an classification problem at the same time on aerial objects:
- Birds (2 types)
- Army aeroplanes (3 types)
- Civil aeroplanes (3 types)
Each of these objects was represented on the image with different backgrounds, which representing different weather condidtions.
For solving this problem i was using modified and deeply simplified version of YOLO architecture.
Civil plane class with probability 0.4:
Fighter plane class with probability 0.47:
Fighter plane class with probability 0.58:
There are 3 independent modules of program. The settings of their work are written in unified config.json.
- Gendata: This module is for generating data for training the neural net. The output data is represented as images and XML annotations. It's written in С++, because the programm is generating data from 3D models, and the simpliest way of manipulating with them on Linux was OpenSceneGraph.
- Archdata: This module is for archiving data and it's written in Python. This module is translates the images and annotations to HDF5 archives with Numpy arrays in the form of dirrect input-ouput of neural net.
- Network: The last but not least. This is the main module. It's written in Python with stack Keras/Theano/cuDNN/CUDA. It just simply has object Network, which you can train or use it for predictions.
Gendata module:
Archdata:
Network: