|"A Brief Recent History of Optical Flow" |
|Speaker ||: ||Dr. Fatma Güney |
|Date ||: ||3 December 2019 (Tuesday) |
|Time ||: ||15:30 - 16:30 |
|Location ||: ||Faculty of Computer and Informatics Engineering, |
İdris Yamantürk Conference Hall (1304)
Traditional approaches addressing the optical flow problem integrate simple local smoothness assumptions about the optical flow field using variational optimization. To overcome the limitations of local priors, sparse matches, and patch-based MRF formulations have been exploited. More recently, deep neural networks have been successfully applied to the optical flow problem. Learning to solve optical flow in an end-to-end fashion from examples is attractive as deep neural networks allow for learning more complex hierarchical flow representations directly from annotated data. However, training such models requires large datasets and obtaining ground truth for real images is challenging. First, by exploiting the power of high-speed video cameras, we propose a method for creating accurate optical flow reference data in a variety of natural scenes. However, occlusions still remain a common source of error. Second, we propose a framework for unsupervised learning of optical flow and occlusions over multiple frames. Our approach based on the guided exclusion of information in occluded regions leads to large improvements and achieves state-of-the-art among unsupervised methods and even a comparable performance to supervised ones. Related papers; 1
Asst. Prof. at Koc University in Istanbul. Previously, postdoc at VGG, at the University of Oxford, working with Andrea Vedaldi and Andrew Zisserman; and Ph.D. student at the MPI for Intelligent Systems, working with Andreas Geiger. Interested in 3D Computer Vision and representation learning from video sequences. Currently working on modeling object-object relations in a video, e.g. multi-object tracking, video object detection, and background motion modeling. Previously worked on action recognition, optical flow estimation, (stereo) depth estimation, and multi-view 3D reconstruction. Her research has been published at top Computer Vision conferences including CVPR, ICCV, and ECCV.