Daily CVPR - Thursday

This project creates adaptive 3D sensors based on information maximization . Instead of treating 3D sensors as a black-box and asking what to do with the 3D data we get or where to move the sensor to learn more about the scene, we are asking: if I already know something about the environment, and I have a specific task in mind, how should the sensor operate? The CSAIL / MIT team focused on structured light scanners, that emit/project light patterns over time and acquire images. The question is: which pattern should the sensor project next? For example, in order to reconstruct a specific object that moves, you may not need to illuminate the whole scene. If you just need to localize in a known environment or help a robot avoid obstacles, you may need radically different patterns than if you want full scene reconstruction for augmented reality. This research paper explores such options based on information maximization and sensor planning. Taking these concepts from decision theory and robotics, it is shown that with the right probabilistic model, they can be used inside the sensor as well. The team chose a probabilistic model that incorporates the scene and sensor pose uncertainty, and yet allows to approximate the information gain between the acquired images and subsets of variables in the scene such as the sensor pose or aspects of the geometry. The model allows to do so in a highly-parallel way, which, it is hoped, will make it useful for real systems. Adaptive 3D scanners, and the concepts shown are expected to result in more efficient and accurate sensors that are better suited to the kind of multiple roles we expects robots and mobile devices will play in the future. 14 Presentations CVPR Daily: Thursday Guy Rosman - CSAIL/MIT