I'm developing a deep learning algorithm to find and classify (post consumption) plastic bottles on a conveyor belt. A way to train my model is to get tons of labeled images of each bottle, which for now is not feasible. Another possible option is to make a 3D scan of the bottle and render it in different angles thus creating synthetic training data.
The way I want to obtain the mesh is by using for example ReMake or CapturingReality and 2D images. The problem is that most of the bottles are transparent and are causing problems in those programs. I just ordered Chalk spray to paint them and with a marker add dots allover. Then hopefully I'd be able to render it with the specific material.
My question is, is there a way to add the visual details like the labels and the color of the bottle cap back to the mesh? I'll be using a RGB-D camera so the priorities are getting both the visual cues as well as the depth information as accurate as possible to what the camera will be classifying in a further stage.
Any advise and feedback will be highly appreciated, thanks in advanced!!