When generating synthetic data, what do you all think provides more value and why? I’ve read research that NNs focus more on textures, but what distinguishes textures from say - a metallic material with randomized color?
Hello @MontyPy! I’ve asked the dev team to help answer your question. Thanks for reaching out to us!
Hello @MontyPy! Update from the dev team:
It will depend heavily on what the downstream task being trained is. I wouldn’t say that texture randomization is necessarily the most important thing in general. For example, if you’re doing a pose estimator network that you want to support a large number of objects, you might be ok to just use a dataset with the appropriate texture for each, and vary lighting conditions, backgrounds, positions, and orientation.