Georgia Institute of Technology

Abstract:

Semi-global matching (SGM) is a widely used and efficient stereo matching algorithm. I will present two improvements to SGM that add very little computational overhead. The first idea is extending SGM to incorporate precomputed surface orientation priors. This improves the algorithm's accuracy on weakly textured slanted surfaces. The second idea is to replace the ad-hoc cost aggregation strategy used in SGM with one that is based on learning. We treat the 1D scanline optimization solutions in SGM as independent disparity proposals and use a random forest to select one of the proposals. We observe consistent improvements over traditional SGM and good generalization performance. I will then describe an efficient multi-frame technique to recovery scene flow from stereo sequences. The method uses SGM to recover disparity and optical flow but also computes camera ego-motion and segments moving objects all within a unified framework.

The last part of the talk will be on learning-based methods for 6-dof object pose estimation in single RGB images. I will describe a new single shot convolutional neural network (CNN)-based architecture for object instance recognition and 6-dof object pose prediction. Our CNN has a simpler architecture and runs significantly faster compared to existing deep 6D object detection methods.

Bio:

Sudipta Sinha is a researcher at Microsoft Research Redmond. His research interests lie in computer vision, robotics and computer graphics. He works on various topics in 3D scene reconstruction from images and video such as structure from motion, visual odometry, stereo matching, optical flow, multi-view stereo, image-based localization, object detection and pose estimation. He is interested in applications such as depth sensing, augmented reality (AR) and UAV-based aerial photogrammetry and 3D scanning. He received his M.S. and Ph.D. from the University of North Carolina at Chapel Hill in 2005 and 2009 respectively. As a member of the UNC-CH team, he received the best demo award at CVPR 2007 for one of the first scalable, real-time, vision-based urban 3D reconstruction systems. He has served as an area chair for 3DV 2016, ICCV 2017 and 3DV 2018, was a program co-chair for 3DV 2017 and serves as an associate editor for the Computer Vision and Image Understanding (CVIU) Journal.

 Website: https://www.microsoft.com/en-us/research/people/sudipsin/