SciViz Panel: Dataset generation for Visual SLAM and Machine Learning
In this talk I will discuss the use of Blender for Applications in Artificial Intelligence (AI). More specifially, I am going to present two of my scientific publications from early 2019 which heavily rely on the use of Blender in order to understand AI technology better. One is about the systematic analysis of the direct Visual SLAM algorithm "Direct Sparse Odometry (DSO)" by generating specific scenarios within Blender. The other one is about evaluating the uncoupled fusion of DSO with a Global Positioning System (GPS) sensor to investigate how to solve one of the main drawbacks with monocular Visual SLAM: Scale ambiguity.
This talk is divided into two main parts. Firstly, I will give a brief overview of what Visual SLAM is and how it can be seperated into two categories: Indirect and direct methods. I will also present the state of the art methods for both approaches: DSO and ORB-SLAM. After discussing the fundamentals I will focus on how this is related to Blender and why we may be interested in not only evaluting the algorithms using Blender but also how direct methods are interesting for artists looking at new ways for detailed 3D reconstruction and matchmoving applications.
The second part will discuss the role of synthetic data in AI. Here, I will focus on object detection and how Blender may help us solving classification problems of objects where labeled groundtruth is either hard to collect or simply not available yet.
The goal of this talk is to motivate the use of Blender in robotics research due to its powerful feature set and open culture. Where possible, this talk will present current research work at our robotics chair at the Friedrich Alexander University in Germany.