南京大学计算机软件新技术国家重点实验室

摘 要：

Traditional computer graphics conducts accurate simulations by explicitly modeling surface geometry, surface reflectance and lighting via the rendering equation. They manage to produce high quality rendering but at an ultra-high computational cost. Conversely, traditional computer vision, in particular 3D reconstruction, seeks to recover camera pose, scene geometry, surface reflectance, etc, from the imagery data via techniques such as correspondence matching and bundle adjustment. In this talk, I present our recent efforts on employing neural modeling and rendering techniques to overcome the limitations in traditional rendering and 3D reconstruction. For graphics I demonstrate deep learning techniques that tackle unknown surface reflectance, corrupted/incomplete 3D shape, and volumetric opacity to produce unprecedented visual quality. For vision, I present a novel volumetric neural reconstruction framework that significantly outperforms state-of-the-art structure-from-motion and photometric stereo methods in reconstruction accuracy. Finally, I discuss how such neural representations may fundamentally change computer vision and graphics and potentially lead to a paradigm shift.