Superpixel based image sequence representation and motion estimation


Tezin Türü: Doktora

Tezin Yürütüldüğü Kurum: Orta Doğu Teknik Üniversitesi, Mühendislik Fakültesi, Elektrik ve Elektronik Mühendisliği Bölümü, Türkiye

Tezin Onay Tarihi: 2017

Öğrenci: KUTALMIŞ GÖKALP İNCE

Danışman: ABDULLAH AYDIN ALATAN

Özet:

In this study a superpixel based representation of image sequences is proposed. For superpixel extraction, a novel gradient ascent approach, in which spatial and spectral statistics are utilized to obtain an optimal Bayesian classifier for pixel to superpixel label assignment, is proposed. Utilization of the spectral and spatial statistics reduce the dependency on user selected global parameters, while increasing the robustness and adaptability. Proposed Local Adaptive Superpixels (LASP) approach exploits hexagonal tiling, while achieving some refinement during initialization in order to improve the computation time and accuracy. The experiments conducted on Berkeley segmentation database show that LASP outperforms the existing methods in terms of boundary recall and computation time. Moreover, the proposed method provides lower bleeding error performance compared to the existing gradient ascent techniques. In order to obtain temporally consistent superpixels, a superpixel based occlusion aware layered motion estimation method is also proposed. Proposed motion estimation method combines a Bayesian method with well known gradient descent approaches for optical flow estimation. Proposed method is able to handle occlusions and large displacements. Experiments conducted on Middlebury Database show that performance of the proposed motion estimation method is comparable to state-of-the-art methods, while providing a less computationally complex solution. Using the output of the motion estimation algorithm, the superpixels in the previous frame placed on the current frame, which provide an initial estimate for superpixels on this frame. Refining this estimate with the information on current frame, it becomes possible to obtain temporally consistent superpixels. These superpixels can be utilized for the representation of image sequences. This representation is developed for video object segmentation, but might also be utilized for various computer vision problems like compression, object tracking and background modeling.