Toolkit to Change the Domain of Computational Imaging Systems
Abstract
Image processing algorithms map an input image into an output image. This output is also known as a target image. Ordinarily, it is assumed that the space of target images is well characterized and/or that examples of target images can be provided. For example, when mapping blurry images to sharp images, one knows what a sharp image should look like and also has a dataset of many examples of blurry and sharp image pairs. Finding the transformation boils down solving for a known-unknown. In contrast to previous work, this proposal considers the case when both the transformation and the target space is unknown. We refer to this problem as Blind Image Translation. If successful, blind image translation is precisely geared toward never-before-seen areas of image processing. After all, the target images have never been observed in previous work. The proposal considers one such example in detail; that of imaging bias. It is widely reported that imaging systems and algorithms are biased against certain demographics. For example, pedestrian detection is more difficult on darker skin tones. The core question we would like to address is whether we can build a camera system that capture ÒcolorlessÓ representations of the world. It is not clear what a colorless representation looks like, which necessitates the blind domain translation framework. If successful, colorless imagesÑ-when used downstream in a computer vision pipelineÑdo not exhibit any performance bias that disadvantages certain skin tone classifications. To fabricate the colorless camera, we develop a method to project the blind domain translation to the set of feasible optical parameters. This enables us to learn a map which can be optically instantiated (through the use of UCLA fabrication facilities). The colorless camera is then used to demonstrate resistance to bias on three visual applications: pedestrian detection, remote plethysmography (rPPG) and a visual Turing Test. Longer-term, the Òspace of fair imagesÓ is only one exemplary case of an unknown target space. The proposal lays a foundation for image processing of input images into unknown target spaces.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 25, 2021
- Source ID
- W911NF2110216
Entities
People
- Achuta Kadambi
Organizations
- Army Contracting Command
- United States Army
- University of California, Los Angeles