IC Colloquium: Sampling and Reconstruction of High-Dimensional Visual Appearance
Event details
| Date | 30.09.2019 |
| Hour | 10:15 › 11:15 |
| Location | |
| Category | Conferences - Seminars |
By: Ravi Ramamoorthi - University of California San Diego
Video of his talk
Abstract:
Many problems in computer graphics and computer vision involve high-dimensional 3D-8D visual datasets. Real-time image synthesis with changing lighting and view is often accomplished by pre-computing the 6D light transport function (2 dimensions each for spatial position, incident lighting and viewing direction). A similar challenge is found in Monte Carlo rendering by sampling the paths of light transport, requiring integration over a 3D-8D space involving time, lens effects for depth of field, pixel area, soft shadows and global illumination. Realistic image synthesis also often involves acquisition of appearance data or light transport from real-world objects, and view synthesis from acquired images, involving 4D-6D functions. In computer vision, problems like lighting insensitive facial recognition similarly involve understanding the space of appearance variation across lighting and view. Since hundreds of samples may be required in each dimension, and the total size is exponential in the dimensionality, brute force acquisition or precomputation is often not even feasible.
In this talk, we describe a signal-processing approach that exploits the coherence, sparsity and inherent low-dimensionality of the visual data, to derive novel efficient sampling and reconstruction algorithms. We describe a variety of new computational methods and applications, from affine wavelet transforms for real-time rendering with area lights, to space-time and space-angle frequency analysis for motion blur and global illumination, to compressive sensing and deep learning for light transport acquisition and view synthesis. The results have had substantial impact, with sampling and denoising now adopted in all production rendering, and we will show recent results for relighting or changing view from only five or six images. The work also points toward a unified sampling theory applicable to many areas of signal processing, computer graphics and computer vision.
Bio:
Ravi Ramamoorthi is the Ronald L. Graham professor of Computer Science at the University of California, San Diego, and founding Director of the UC San Diego Center for Visual Computing. Prof. Ramamoorthi is an author of more than 150 refereed publications in computer graphics and computer vision, including 75 at ACM SIGGRAPH/TOG, and has played a key role in building multi-faculty research groups that have been recognized as leaders in computer graphics and computer vision at Columbia, Berkeley and UCSD. His research has been recognized with a half-dozen early career awards, including the ACM SIGGRAPH Significant New Researcher Award in computer graphics in 2007, and the Presidential Early Career Award for Scientists and Engineers (PECASE) for his work in physics-based computer vision in 2008. He was elevated to IEEE and ACM Fellow in 2017, and inducted into the SIGGRAPH Academy in 2019.
More information
Video of his talk
Abstract:
Many problems in computer graphics and computer vision involve high-dimensional 3D-8D visual datasets. Real-time image synthesis with changing lighting and view is often accomplished by pre-computing the 6D light transport function (2 dimensions each for spatial position, incident lighting and viewing direction). A similar challenge is found in Monte Carlo rendering by sampling the paths of light transport, requiring integration over a 3D-8D space involving time, lens effects for depth of field, pixel area, soft shadows and global illumination. Realistic image synthesis also often involves acquisition of appearance data or light transport from real-world objects, and view synthesis from acquired images, involving 4D-6D functions. In computer vision, problems like lighting insensitive facial recognition similarly involve understanding the space of appearance variation across lighting and view. Since hundreds of samples may be required in each dimension, and the total size is exponential in the dimensionality, brute force acquisition or precomputation is often not even feasible.
In this talk, we describe a signal-processing approach that exploits the coherence, sparsity and inherent low-dimensionality of the visual data, to derive novel efficient sampling and reconstruction algorithms. We describe a variety of new computational methods and applications, from affine wavelet transforms for real-time rendering with area lights, to space-time and space-angle frequency analysis for motion blur and global illumination, to compressive sensing and deep learning for light transport acquisition and view synthesis. The results have had substantial impact, with sampling and denoising now adopted in all production rendering, and we will show recent results for relighting or changing view from only five or six images. The work also points toward a unified sampling theory applicable to many areas of signal processing, computer graphics and computer vision.
Bio:
Ravi Ramamoorthi is the Ronald L. Graham professor of Computer Science at the University of California, San Diego, and founding Director of the UC San Diego Center for Visual Computing. Prof. Ramamoorthi is an author of more than 150 refereed publications in computer graphics and computer vision, including 75 at ACM SIGGRAPH/TOG, and has played a key role in building multi-faculty research groups that have been recognized as leaders in computer graphics and computer vision at Columbia, Berkeley and UCSD. His research has been recognized with a half-dozen early career awards, including the ACM SIGGRAPH Significant New Researcher Award in computer graphics in 2007, and the Presidential Early Career Award for Scientists and Engineers (PECASE) for his work in physics-based computer vision in 2008. He was elevated to IEEE and ACM Fellow in 2017, and inducted into the SIGGRAPH Academy in 2019.
More information
Practical information
- General public
- Free
- This event is internal
Contact
- Host: Wenzel Jakob