Gold

Friends

Chair: Andres Bruhn

Special Invited Talk

Low-level Vision for Autonomous Driving

Raquel Urtasun
TTI Chicago

Slides

Over the past few decades we have witnessed a great improvement in performance of low-level vision algorithms. As demonstrated by the Middlebury leader-board, problems such as stereo and optical flow estimation are mostly solved when dealing with imagery captured in the laboratory. Unfortunately, when dealing with real world problems most algorithms fail.

In this talk I’ll show how autonomous driving can provide us with very challenging scenarios to push current algorithms to the next level. Furthermore, I‘ll argue that the next generation of low-level vision algorithms should be holistic and reason jointly about multiple related tasks such as recognition and perceptual grouping.
An Evaluation of Data Cost Functions for Optical Flow Estimation
Christoph Vogel, Konrad Schindler, and Stefan Roth
Illumination Robust Optical Flow Model Based on Histogram of Oriented Gradients
Hatem Rashwan, Mahmoud Mohamed, Miguel Angel, Bärbel Mertsching, and Domenec Puig