Deep Superpixel Stereo

Author:

Sun, Qian

Graduate Program:

Information Sciences and Technology

Degree:

Master of Science

Document Type:

Master Thesis

Date of Defense:

April 01, 2019

Committee Members:

Zihan Zhou, Thesis Advisor/Co-Advisor
James Z Wang, Committee Member
Amulya Yadav, Committee Member

Keywords:

Stereo Matching
Superpixel
Deep Neural Network

Abstract:

Disparity estimation is an essential problem in computer vision. Recent work has shown that disparity estimation can be formulated as an end-to-end supervised learning task and achieves impressive performance. However, existing methods may still fail in textureless areas and occluded regions. To alleviate such challenges and improve the disparity estimation results, in this thesis we propose to incorporate a slanted-plane model into existing deep learning frameworks. We propose a novel end-to-end trainable deep neural network called Deep Superpixel Stereo Network, which can simultaneously generate superpixels and predict the plane parameters of each superpixel. Our network is developed based on DispNetC, a popular and effective network architecture for stereo matching. Experimental analysis shows that, with the integration of superpixels and the parametric model, our method outperforms the original DispNetC on the FlyingThings3D dataset and the KITTI 2015 dataset. Further, our method is able to generate superpixels which adhere to object boundaries well and are also compact. Experiment results demonstrate that our method achieves competitive performance against the state-of-the-art methods in terms of superpixel segmentation. Our work suggests that the integration of superpixels could potentially result in performance improvement in other dense prediction problems in computer vision, such as depth prediction and optical flow estimation.