Zpy is an open-source platform for creating synthetic data in Python. The tool is developed on the basis of Blender for use in computer vision tasks.
Synthetic data generation is actively used in computer vision tasks due to the need to use massive training datasets, which are often difficult or impossible to collect based on real data. You can work with zpy using three interfaces: the Python API, the command line, and a graphical web application. Zpy allows you to randomize domains, that is, to increase the degree of data diversity. As applied to images, domain randomization can be implemented by changing the illumination of each image or the textures of objects. The figure below shows the results of applying the three randomization methods supported by zpy (from top to bottom):
changing the background on the image;
changing the position and number of light sources;
changing the texture.
The last panel shows the result of applying all the randomization methods.
ZPY also provides the possibility of automatic semantic segmentation of pixels. Example of segmentation using zpy:
The zpy code is available on GitHub.
