I'm working on an app to identify LEGO elements from a photo using machine learning. I've completed a crude prototype that only supports the most common 10 elements.
I'm currently rendering images for the dataset using LeoCAD, but this isn't ideal, as I'd like the training data to look as much like real-world photos as possible to improve accuracy. I'm aware that I could get better renders using Blender or other tools, but that wouldn't provide all the elements of real photos (odd backgrounds, shadows, poor lighting, etc).
Is there an open dataset of LEGO part photos suitable for machine learning applications? I'm aware of a Kaggle dataset, but it uses renders rather than photos. I'm also aware of a couple of folks (Daniel West and Jacques Mattheij) who have used a dataset like the one I'm looking for, but I believe they created it themselves and did not share it for public use.
I've heard from both Daniel and Jacques. Daniel's dataset has already been exclusively licensed. Jacques wouldn't be easy to generalize due the optics used (mirrors to capture multiple angles), and has already been deleted.