The reconstruction of the 3D world from images is among the central challenges in computer vision. Starting in the 2000s, researchers have pioneered algorithms which can reconstruct camera motion and sparse feature-points in real-time. In my talk, I will show that one can autonomously fly quadrotors and reconstruct their environment using onboard color or RGB-D cameras. In particular, I will introduce dense and direct methods for camera tracking and reconstruction which do not rely on keypoint extraction, which exploit all available input data and which recover dense geometry rather than sparse point clouds. This is joint work with Jakob Engel, Vladyslav Usenko, Jan Stuehmer, Martin R. Oswald, Christian Kerl, Erik Bylow, Jörg Stückler and Juergen Sturm.

This talk will introduce our recent deep learning works in video surveillance. The applications of deep learning include object detection, pedestrian detection, person re-identification, general object tracking, crowd segmentation, crowd density estimation, crowd counting and crowd video classification. Many results have shown that deep learning can advance the state-of-the-art of video surveillance substantially. The focus of this talk would be the strategies of designing network structures and learning feature representations adapting to surveillance applications. With carefully designed network structures and training schemes, the learned features could be effective for general objects, a particular object class, a particular object instance, or a large group of people to fulfill the requirements of different surveillance applications. In video surveillance, it is also critical for the feature representations to be robust across a large number of diversified scenes and camera views, to be robust to background clutters, and to well motion information. Deep learning is effective on addressing these challenges.

Coming soon