Visual Object Recognition and Scene Interpretation
Personal Homepage: http://www.d2.mpi-inf.mpg.de/People/stark
The interpretation of visual scenes through computer vision algorithms has gained increasing importance in a wide variety of applications, such as information retrieval, robotics, and autonomous driving.
A major challenge lies in the severely under-constrained nature of the scene interpretation problem, since visual observations alone are inherently ambiguous – one and the same image could possibly have been generated by a multitude of different physical scenes. Equally, the same physical scene might give rise to different semantic interpretations, depending on contextual information available to the observer.
In our research, we aim to meet this challenge by following and combining two orthogonal directions. The first direction focuses on the robust recognition of individual objects that together constitute a scene, imposing an initial set of constraints on the space of plausible scene interpretations. For that purpose, we intend to extend learning-based object class recognition techniques to provide accurate estimates of object pose, viewpoint, and three-dimensional extent.
The second direction reasons about a scene in its entirety, drawing from prior knowledge about feasible scene geometry and interactions among multiple objects, which further constrains the space of plausible interpretations. The identification of maxima in this space will require the design of efficient inference procedures, which will be an important focus of our research.
Since 2015 Michael Stark is a scientist at Vicarious, an A.I. company in San Francisco.