In this paper, we propose a reconstruction technique that uses 2D regions/superpixels rather than point features. We use pre-segmented RGBD data as input and obtain piecewise planar 3D models of the world. We solve the problem of superpixel labeling within single and multiple views simultaneously by using a Rao-Blackwellized Markov Chain Monte Carlo (MCMC) algorithm. We present our output as a labeled 3D model of the world by integrating out over all possible 3D planes in a fully Bayesian fashion. We present our results on the new SUN3D dataset [?].

No abstract

Robot navigation already has many relatively efficient solutions: reactive control, simultaneous localization and mapping (SLAM), Rapidly-Exploring Random Trees (RRTs), etc. But many primates possess an additional inherent spatial reasoning capability: mental rotation. Our research addresses the question of what role, if any, mental rotations can play in enhancing existing robot navigational capabilities. To answer this question we explore the use of optical flow as a basis for extracting abstract representations of the world, comparing these representations with a goal state of similar format and then iteratively providing a control signal to a robot to allow it to move in a direction consistent with achieving that goal state. We study a range of transformation methods to implement the mental rotation component of the architecture, including correlation and matching based on cognitive studies. We also include a discussion of how mental rotations may play a key role in understanding spatial advice giving, particularly from other members of the species, whether in map-based format, gestures, or other means of communication. Results to date are presented on our robotic platform.

No abstract

Abstract. In this paper, we present an algorithm for doing high frame rate egomotion estimation. We achieve this by using a basis flow model, along with a novel inference algorithm, that uses spatio-temporal gradients, foregoing the computation of the slow and noisy optical flow. The inherent linearity in our model allows us to achieve fine grained parallelism. We demonstrate this by running our algorithm on GPUs to achieve egomotion estimation at 120Hz.Image motion is tightly coupled with the camera egomotion and depth of the scene. Hence, we validate our approach by using the egomotion estimate to compute the depth of a static scene. Our applications are aimed towards autonomous navigation scenarios where, it is required to have a quick estimate of the state of the vehicle, while freeing up computation time for higher level vision tasks.