Synthetic Data Generation for Visual Detection of Flattened PET Bottles

Polyethylene terephthalate (PET) bottle recycling is a highly automated task; however, manual quality control required due to inefficiencies of the process. In this paper, we explore automation sub-task, namely visual detection, using convolutional neural network (CNN)-based methods and synthetic generation labelled training data. We propose pipeline tailored for transparent crushed PET detection; it can also be applied undeformed bottles if viewpoint set from above. conduct various experiments on CNNs compare real data, show that data reduce amount experiment with combination both datasets in multiple ways obtain best performance.