Machine learning LEGO image recognition: Using virtual data and YOLOv3

I have been working a lot with LEGO and 3D models lately. For my current project I am looking to build a LEGO image recognition program. My ideal scenario is to grab a handful of LEGO, toss them on the table, take a picture, and have the program catalog the pieces. The biggest challenge I encounter with any machine learning project is collecting and formatting the training data. I am pretty sure this is the biggest challenge everyone encounters with machine learning.