BRADLEY KOSKOWICH, GSCS PH.D. STUDENT, defends phd proposal
On 2/13/20, BradleyKoskowich of MANTIS Lab passed his PhD qualifying exams on his proposed topic, Efficiently Localizing Monocular Images Using Image Synthesis, Point Cloud Products & Keypoint Densification. See his abstract below for more details!
ABSTRACT:
Human beings possess powerful visual abilities which are used to navigate space, coordinate our actions, and correlate the contents of a perspective image with an approximate knowledge of the location the image would have been taken from, intuitively and from little more than the image itself. This lends itself to further abilities to create logical conclusions about the nature of an inferred three-dimensional space based on a series of two-dimensional observations. Recreating this same process digitally is a thoroughly studied problem which has seen years of computer vision research, namely feature detection & description, structure-from-motion (SfM), and their applications in simultaneous localization and mapping (SLAM). Variations and permutations of SLAM solutions have been proposed since its initial inception, and as of recent years have leaned towards incorporating additional hardware to improve their results, such as inertial motion units and global navigation satellite systems. However, the nature of incorporating additional data dimensions in these SLAM solutions alters it such that it the problem becomes intractable in the event of catastrophic interference or data loss from any dimension. While there have been attempts to mitigate or even read hardware interference and leverage it as additional information, such methods are not robust to even a partial systems failure. And systems which have proven robust to such failures incur significant pre-processing or computational time overhead, none of which propose to be suitable for real-time applications. This proposal assesses the need to decouple the localization component of SLAM solutions from the rest of the processing pipeline, allowing for flexible input datasets with tolerance for extreme noise. This proposal details mechanisms which can be developed and combined specifically to enable high performance POSE estimation as part of this flexible localization process, later incorporated into a robust and tolerant SLAM solution as a hybrid of typical algorithms and neural networks.