Program to generate images from a 3D scene with objects. Detects each object and randomly sets the display value. This value is stored in a '1-hot-vector' for use as feedback in a convolutional neural network. This project is meant to showcase how useful Blender can be to generate synthetic data for testing neural networks.
- Blender 2.78
- Python 2 and 3
- OpenCV or PIL
See SyntheticData_28_7_2017.pdf for an extensive explanation behind the concept of synthetic generation for training neural networks.
First check basic_shape_gen.blend to inspect or modify the actual contents of the scene. Then modify and run render.sh to render an imagedata set and generate a .csv reference file. When rendering is complete, check out the 'model' directory, which contains the convolutional model and an input data processing script.
Check the filepaths so that you can easily input the render output into the training program.
An interesting test is to add different objects to the scene, and animate more parameters to determine wether this basic network is able to distinct the same objects.
See example files labels_files.csv and labels_files_2.csv on what the render program outputs. Changing the objects in the scene, and adding 'ai_' to the object name will have it included by the render script in the labels and randomized display.
See LICENCE