Incremental Learning for Optical Flow

Tomislav Dobricki, Yoonghee Oh, Haeun Ko, Taeyun Kim, Dongyooung Kim, Byung-Woo Hong

Abstract: This paper presents a potential application of deep learning algorithms to the computation of optical flow in a temporal sequence of images. The motion information obtained by optical flow provides significant information regarding the analysis of video. The computational challenge for the estimation of optical flow makes it difficult for real-time applications. Thus, we explore a method that has a potential in improving computational efficiency for optical flow, in the framework of deep learning, by incorporating a learning technique that takes into consideration incremental warping. The presented paper is limited in that numerical results are not provided, yet it gives a direction for further studies.