서강대학교 전자공학과

일시 및 장소: 2019년 9월20일 10:30, RA204

강사: 임성훈 교수, DGIST 정보통신융합전공

제목: Incorporating Geometric Knowledge into a Deep Neural Network 
부제:  - to Explore Synergies in 3D Computer Vision -

초록: Recently, many smartphone manufacturers are actively studying 3D reconstruction to enable various AR / VR and photography applications. To support the features, they have developed a new multi-camera system mounted on mobile phones and it has become a new trend in the mobile industry. However, the performance of depth estimation heavily depends on the scene environment, such as texture-less region and illumination changes. To address these photometric issues, I proposed a robust 3D reconstruction method for such challenging conditions. I achieved this by investigating the advantages of two different research field, traditional 3D geometry and recent breakthrough in deep learning. I mainly studied what knowledge of 3D geometry to incorporate into a deep neural network, and how to integrate them to explore synergies in 3D computer vision. In this talk, I will share my two different attempts and results of the incorporation. Part of the content of this talk was published in CVPR2018 and ICLR2019.