Bringing Images to Light

点击翻译按钮获取中文摘要

Practical computer vision applications have, for a long time, found themselves confined to the realm of "machine vision": the image-based technology for automatic inspection and analysis used on assembly lines. In recent years, and in combination with the advent of cheap digital cameras, we have witnessed an emergence of computer vision applications in massive-scale consumer applications. For example, "intelligent" cameras automatically detect faces and shoot when the person is smiling, search engines successfully retrieve images from a text query, online digital maps are augmented with street-level imagery, and augmented reality applications are commonly available on smartphones. Despite all these advances, this is only the beginning. The computer vision research community is actively working on much more challenging tasks, such as the fully automated understanding of images, which are bound to find their ways onto commercial progress in the not-so-distant future. **Despite this progress, computer vision systems suffer from one major limitation: they have trouble adapting to the strongly-varying illumination conditions, especially outdoors. In particular, the sun, with its blinding intensity, creates effects such as highlights and cast shadows that are very difficult to deal with. While the human eye can easily adapt to those variations through millions of years of adaptation, the same cannot be said of digital cameras and of algorithms operating on their images. **This research program will focus on the issue of allowing computer vision algorithms to adapt, and even leverage, the lighting conditions in an image. We will develop the new paradigm that jointly reasoning about illumination and scene components can help us better understand images. To do so, we will tackle the following four objectives. We will: 1) introduce new inference algorithms for jointly reasoning about face appearance and illumination conditions, to improve face detection in challenging lighting conditions; 2) develop new lighting models to capture the appearance of objects (people and cars) in a data-driven way to build better object detectors; 3) incorporate physical models of appearance to better segment images into semantic regions; and 4) provide a general framework for understanding how illumination is a useful cue in computer vision, even in the challenging world of consumer photographs. **In addition to the applications mentioned above, the activities proposed in the context of this research program will also impact other fields, such as computer graphics, intelligent transportation, and robotics, all relevant to Canada. In computer graphics, additional information on the lighting conditions in images can contribute to improving applications such as pre-visualization, digital asset capture and management, and image editing. Safety features on intelligent transportation systems such as autonomous cars, which often rely on cameras for detecting obstacles on the road, strongly depend on their capacity to adapt to varying illumination conditions. Similarly, outdoor exploratory robots will also benefit from this research through the development of better models of outdoor illumination. The creation of this research program at Laval University will contribute to maintain and even expand Canada's relevance in these fields via its key technical contributions, and by training highly qualified personnel in its development.