Prof. Fisher Yu

Computer Vision Lab, D-ITET,
ETH Zurich, Switzerland

Visual Motion Understanding for Autonomous Driving

Understanding motion from visual inputs is a critical task for driving a vehicle in a complicated dynamic environment. The task can be divided into three levels: scene, object, and pixels. In this talk, I will start with image-level scene prediction conditioned on actions and visual history and introduce our Semantic Predictive Control framework. The predicted semantic segmentation of future frames can provide efficient guidance for end-to-end driving policy learning. To shed light on how we can future improve the policy learning robustness, I will dive into the object-level motion analysis to show that we can learn 3D vehicle motion from only monocular videos. A key component for visual motion analysis is to find correspondences across images. Therefore, I will discuss our Hierarchical Discrete Distribution Decomposition (HD^3) method. The method will provide not only accurate point-wise correspondence estimation, but also reliable uncertainty and distribution information. I will conclude the talk with certain future directions on algorithms and datasets for motion analysis. Prof. Fisher Yu will be joining as Tenure Track Assistant Professor of Computer Vision at ETH Zurich. Fisher Yu conducts research on computer vision and machine learning. His work covers a wide spectrum, ranging from the basics of machine image and video analysis through to practical applications, such as in self-driving vehicles. His principal tools are neural networks, which he both applies and develops further. He strives to achieve the most universal representation of visual understanding. Fisher Yu's appointment will help ETH Zurich and the Department of Information Technology and Electrical Engineering to safeguard their leading role in computer vision and strengthen them in the area of machine learning.