*******************************************************************
KOÇ UNIVERSITY
GRADUATE SCHOOL OF SCIENCES & ENGINEERING
COMPUTER SCIENCE AND ENGINEERING
MS THESIS DEFENSE BY EKREM EMRE YURDAKUL
******************************************************************
Title: Semantic Segmentation of RGBD Videos with Recurrent Fully Convolutional Neural Networks
Speaker: Ekrem Emre Yurdakul
Time: September 29, 2017, 14:00
Place: ENG 208
Koç University
Rumeli Feneri Yolu
Sariyer, Istanbul
Thesis Committee Members:
Assoc. Prof. Yücel Yemez (Advisor, Koc University)
Prof. Dr. A. Murat Tekalp (Koc University)
Prof. Dr. Gözde Ünal (Istanbul Technical University)
Abstract:
Semantic segmentation of videos using neural networks is currently a popular task, however the work done in this field is mostly on RGB videos. The main reason for this is the lack of large RGBD video datasets, annotated with ground truth information at the pixel level. In this work, we use a synthetic and a real RGBD video dataset to investigate the contribution of depth and temporal information to the video segmentation task using fully convolutional and recurrent fully convolutional neural network architectures. Additionally, we employ weight transfer from fully convolutional neural networks to recurrent fully convolutional neural networks and investigate different depth encoding schemes. Our experiments show that the addition of depth information improves semantic segmentation results and exploiting temporal information results in higher quality output segmentations.