Occlusion contour (OC) is well known to play important roles in many computer vision tasks. Unlike regular photographs, an OC image removes the effects of illumination, texture, and appearance while preserving important depth edges and silhouette. This project develops a comprehensive framework for acquiring, processing, and utilizing OCs in visual inference tasks. On the sensor front, the research team develops a new Occlusion Contour Camera or OC-Cam. The new OC-Cam extends the multi-flash camera by coupling an array of controllable infrared (IR) LEDs and a visible-IR camera pair. On the algorithm and application fronts, the research team systematically develops OC-assisted visual inference algorithms. For recognition, the acquired OCs are used as a feature filter to improve category-level object recognition. For tracking, the PIs apply OCs to enhance target representation by filtering out the background and texture edges. Furthermore, the research team investigates the previously under-explored problems of OC-assisted image summarization and privacy protection.
This project can cast deep impact on broad areas of computer vision, artificial intelligence, criminal justices, and robotics, both in research and education. Due to the importance of OCs in human vision, the results can produce a testbed for the study of visual psychology. Furthermore, the OC-Cam is expected to serve as conceptual inspiration for constructing the next-generation surveillance systems. Finally, the captured OC datasets and relevant tools are made available to other researchers, to provide a platform for validating new OC-based computer vision algorithms.
