Explore the cutting-edge Large Mask Inpainting (LaMa) system for removing foreground objects from images, even when they cover substantial portions of the image. Learn about the innovative use of Fourier Convolutions to incorporate global information throughout the forward propagation, resulting in superior reconstruction of periodic structures with long-range consistency. Delve into the model architecture, Fourier convolution layers, loss function, and mask generation algorithm. Examine experimental results and gain insights from an exclusive interview with the paper's authors. Discover how LaMa overcomes challenges in inpainting large missing areas, complex geometric structures, and high-resolution images, outperforming previous state-of-the-art methods.
Resolution-Robust Large Mask Inpainting with Fourier Convolutions - With Author Interview
Overview
Syllabus
- Intro
- Sponsor: ClearML
- Inpainting Examples
- Live Demo
- Locality as a weakness of convolutions
- Using Fourier Transforms for global information
- Model architecture overview
- Fourier convolution layer
- Loss function
- Mask generation algorithm
- Experimental results
- Interview with the authors
Taught by
Yannic Kilcher
