Scripts to produce chess board image datasets for classification experiments
The script and blender files were tested using Blender 2.78c and Python 3.6.2. Earlier versions should work as well.
It seems that blender has some problems if new objects are constantly created and removed and new images are rendered. (i.e it becomes much slower over time). Since I was not able to solve this I hacked up a temporary fix by writing an additional python script that repeatetly calls blender to render a smaller batch of images. A batch size of 50 to 100 usually works well.
The base image path is hard coded in the script run_chess_single.py
so you should change it.
Example usage
./run_chess_single.py --start 0 --stop 30000 --batchsize 100 --diff 1 5 10
This will produce 30000 images per class in batches of 100 (blender will be restarted after every batch) for the given problem and for 1, 5, and 10 out of place pawns. So 180000 images will be produces all in all.
If images of a different resolution are needed, the .blend file can
be opened and appropriate changes to the render settings can be made.
It is also easy to change the color and shape of the pawn, different lighting etc. if you are familiar with Blender.