ETH Zürich - Robotics, Systems and Control MSc Semester Project
This semester project tackled the exploration of a database of parametric textures. Its main objective was to generate sufficiently many different combinations of the textures’ parameters (referred to as variants), to be later used as training data for a supervised deep learning method tasked with recognising textures in real images. To this end, a database of textures provided under the Substance format was analysed in terms of the distribution of output channels (texture layers) and editable parameters for each texture.
The set of these channels and parameters was then narrowed down in a principled manner, favouring those that produced variants closest to what one would expect a texture recognition algorithm to encounter. After generating the variants, the pair-wise Mean Squared Error (MSE) was employed as an appropriate metric by which redundant, near-identical variants could be eliminated. Lastly, through random forest regression, the relationship between variants’ parameters and their MSE was learned, creating a model that can prune a set of generated variants to a concise form, fit for training a deep learning model.