3D Synthetic Data for Computer Vision

This project uses a 3D model to train a computer vision algorithm to be able to "see" its 3d printed counterpart in the real world. For more context, here is a video of it in action.

It uses a method of transfer learning, or the process of training an algorithm on top of a model that has been trained on real world data.

ML Model: Yolact
Digital Sculpture Artist: Ben Miller
Immersive Limit Blog: immersivelimit.com

The goal was to train the computer vision model to recognize objects and create masks around them.

This is a high resolution render of the digital 3d model.

3d print of the model in the real world.

Creating the image dataset required rendering several variations from different angles.

With synthetic datasets we can easily create masks for our objects, to feed to the machine learning model.

Example of the trained model creating instance segmentations on real world images.

Real world images were used for backgrounds in the training images

Several thousand training images were created by randomly pasting 3d renders of the digital sculpture over the real world background images.

Screenshot of a 3d viewport that shows the path the 3d camera used to take images of the model.