3D Virtual Environment

We present below a sample dataset collected using our framework for synthetic data collection that is efficient in terms of time taken to collect and annotate data, and which makes use of free and open source software tools and 3D assets. Our approach provides a large number of systematic variations in synthetic image generation parameters. The approach is highly effective, resulting in a deep learning model with a top-1 accuracy of 72% on the ObjectNet data, which is a new state-of-the-art result.