Generating ground truth datasets for virtual reality experiences

Abstract

Systems and methods of generating ground truth datasets for producing virtual reality (VR) experiences, for testing simulated sensor configurations, and for training machine-learning algorithms. In one example, a recording device with one or more cameras and one or more inertial measurement units captures images and motion data along a real path through a physical environment. A SLAM application uses the captured data to calculate the trajectory of the recording device. A polynomial interpolation module uses Chebyshev polynomials to generate a continuous time trajectory (CTT) function. The method includes identifying a virtual environment and assembling a simulated sensor configuration, such as a VR headset. Using the CTT function, the method includes generating a ground truth output dataset that represents the simulated sensor configuration in motion along a virtual path through the virtual environment. The virtual path is closely correlated with the motion along the real path as captured by the recording device. Accordingly, the output dataset produces a realistic and life-like VR experience. In addition, the methods described can be used to generate multiple output datasets, at various sample rates, which are useful for training the machine-learning algorithms which are part of many VR systems.