Taming the Complexity of Light Transport ---------------------------------------- Underlying most of computer vision research is a model of how light interacts with a scene and then reaches a camera to form images. Light propagates through a scene in complex ways: inter-reflections between scene points, diffusion beneath the surface of translucent materials like skin and marble, and scattering through media like the atmosphere and murky water. Despite this complexity, the vision community has historically defined the brightness of a pixel in the image as solely due to the light reflected from a single point in the world. Modeling this wide variety of optical phenomena is crucial for effective scene understanding in all real-world environments. This talk will present computational imaging and illumination techniques to control and tame the complexity of light transport for many applications in the areas of intelligent transportation, long-range imaging, underwater photography, 3D displays and digital effects. Light and Water in Nature ------------------------- The visual manifestations of water in our every day lives are wide ranging. Light interacts with tiny droplets that form mist, fog and clouds, with larger drops of rain and ice crystals of snow flakes, and propagates through large water bodies like oceans, in fascinating ways. Finite element methods are used to model aggregate light transport within volumes of fog, haze and water bodies. On the other hand, light refraction and reflection are modeled through individual particles such as rain drops, snow flakes, and hail stones. This talk will explore these interactions and present computational imaging and illumination tools to enable several applications: visibility improvement in murky waters, photo-realistic rendering of liquids, tracking of objects within water, building a 3D surround display made entirely of water drops, and designing vehicular headlights that see through rain, hail and snow. Light and Water Drops --------------------- Water drops are perfect little lenses. They absorb very little but refract/reflect incident light at virtually all angles. By precisely controlling the positions and timings of water drops, I will show it is possible to create a novel three-dimensional and surround display. Since the display does not have sharp moving components (like mirrors in other 3D displays), they are touchable and can provide new interactive experiences. We are planning to build a large scale (room-sized) display where the consumer can walk right in to and surround himself with the display. Sometimes, however, water drops can be a nuisance. Imagine driving at night in a rainstorm. The bright flickering rain streaks can be distracting to the driver, making driving stressful and dangerous. But if we are able to design headlights that selectively dis-illuminate enough number of rain drops, the visibility of precipitation can be reduced. I will describe the significant in-roads we have made into this challenging project. Finally, I will conclude with my thoughts on using drops as lenses for flexible imaging systems. Imaging and illumination in poor-visibility environments -------------------------------------------------------- Bad weather, murky water, dust, turbulence and smoke obscure our vision making it hard to navigate. Ordinary images captured in these conditions show significant loss of contrast and blur. Our lab is developing computational imaging and illumination tools as well as light transport models and algorithms to increase visibility in such conditions. This talk will summarize the main results in reducing effects of fog, haze, rain, snow, turbulence and murky water in images. Applications of these works include long distance imaging and terrestrial and underwater surveillance and security. As a byproduct of our algorithms, we show that it is possible to automatically estimate the three dimensional structure of the observed scene. Lastly, this talk will show how to add these effects realistically and efficiently to video games and movies. Revisiting 3D reconstruction for complex objects and materials -------------------------------------------------------------- Computing 3D shapes of objects from imagery has been one of the core research problems in computer vision. Significant strides have been achieved in this area as demonstrated by the commercially available stereo systems, laser range finders, and by more recent breakthroughs in reconstruction of large scale scenes like cities. That said, many challenges remain since traditional approaches fail when applied to objects made of complex material properties and shapes. This talk summarizes the research in our group focused on analyzing and exploiting light transport to develop new methods for reconstructing challenging scenes such as shiny surfaces, marble objects, tree branches and corals, as well as on high-speed motion capture. What is the space of all object deformations? --------------------------------------------- Many objects we interact with daily such as the clothing we wear, the wavy water surface, pages of a book, undergo a complex set of deformations. Visual understanding of such deformations can be crucial for a robot to fold clothing or a news paper, an OCR scanner to digitize books, for a camera system to raise alarms if somebody is drowning in a swimming pool or the ocean, or for a designer to virtually visualize clothing on a model. This talk summarizes the results of our research program to model, analyze, and estimate deformations of real-world objects. At the heart of the research is a principled combination of physically-based and data-driven approaches that guarantees globally optimal solutions to hard non-linear estimation problems that are necessary to shed the light on object deformations.