RI: Small: Tasking on Natural Image Statistics: 2D and 3D Object and Category Detection in the Wild

点击翻译按钮获取中文摘要

This project develops "distortion-aware" computer vision models and algorithms suitable for today's mobile camera devices, such as are found in cell phones. Today's mobile camera devices contain remarkably powerful computing capability, sufficient, in fact to contemplate performing sophisticated computer vision problems such as three-dimensional depth estimation, object detection and object recognition. However, mobile camera capabilities are much more limited due to distortions on capture, such as low-light noise, blur, saturation, over/under exposure, and processing artifacts such as compression. These distortions cause most computer vision algorithms to "break," making them unable to accurately recreate the 3D world or to find and recognize objects in it. This project creates computer vision algorithms with similar capability, using new and emerging models of visual neuroscience (how people see) and detailed and accurate statistical models of the three dimensional visual world (called natural scene statistic models). The project can impact many other camera devices, including low-cost surveillance and security cameras, mobile medical cameras, military cameras operating under battlefield conditions, and more.This research develops principled approaches to using natural scene statistics models to solve difficult single-image visual tasking problems under poor imaging conditions. Specifically, the research team studies robust 'distortion-aware' statistical image models and algorithms for single-image 2D and 3D object and object category detection and synergistic 3D depth estimation. The research work includes (1) developing algorithms for fast, generic object detection and categorization "in-the-wild" that operate on single photographic images suffering authentic artifacts from digital cameras; (2) designing object and object class detection mechanisms augmented by 3D depth estimation processes, driven by powerful 2D and 3D prior natural image constraints; and (3) constructing a new annotated Color+3D database of HD precision-calibrated RGBD data using a Reigl VZ-400 Terrestrial Lidar Scanner on object categories of interest, yielding data of higher resolution and richness than existing datasets, complete with image labels as well as hand annotations of bounding-box object locations. This database is free to the community at large once it is available.