Nachricht

Am Donnerstag, den 4. Oktober 2018 hält Herr Emrah Caglar um 16:30 Uhr im Raum P 1.4.18 einen Vortrag über seine Masterarbeit mit dem Titel:

Performance Analysis of Optical Flow Constraints for Unsupervised Learning of Convolutional Neural Networks

Abstract:

Recent studies have shown that the estimation of optical flow, which is one of the fundamental tasks of computer vision, can be solved through deep learning methods. Particularly in the last years, supervised and unsupervised deep learning techniques have been becoming very popular in this field. The performance of supervised deep learning methods for optical flow estimation has shown that they are highly competitive compared to traditional energy-based flow estimation approaches. However, the difficulty of generating ground-truth training data has a negative impact on the accuracy of the optical flow estimation using the supervised learning approaches. In contrast, unsupervised learning techniques solve the above-described training data problem and deliver competing performances akin to supervised variants. The analysis of these unsupervised methods for the flow estimation is crucial. Hence, in this work, unsupervised deep learning methods for the optical flow estimation with different flow constraint assumptions should be studied and analyzed. To solve this task, classical optical flow constraint assumptions will be embedded in the deep learning networks as loss functions to form an unsupervised training for the flow estimation. The performance of the modified networks will then be evaluated and compared based on optical flow datasets.