Real versus Realistically Rendered Ground Truth: Can we use Computer Graphics for Performance Evaluation?

  • Daniel Kondermann | Heidelberg Collaboratory for Image Processing (HCI)

Creating ground truth for optical flow in natural outdoor environments seems almost impossible. In this talk, I will propose two approaches we are currently investigating.
The first approach is to use semi-automatic vision algorithms as is done for example in movie postproduction to create “weak” ground truth. The second approach is to evaluate the properties of today’s computer graphic rendering systems with respect to their ability to generate images close to the real world.
Finally I will discuss the problem of defining performance measures and benchmarking with respect to correspondence estimation and related algorithms.

    • Portrait of Jeff Running

      Jeff Running

Watch Next