Abstract

Cognitive Vision Systems aim to model visual cognitive abilities of humans, for example object detection and recognition, visual attention, visual search, scene recognition etc. In such tasks, humans still outperform current computer vision systems considerably, especially in their capabilities to generalize and to deal with noise, clutter, and unexpected situations. Therefore, a promising approach to improve machine vision systems is to analyze and simulate human cognitive abilities. In this talk, I will present an overview of my recent research on cognitive systems, focusing on object discovery. Object discovery is the task to find unknown objects in a scene. Since the system starts without prior knowledge about the types of objects, it addresses the question "what is an object?". While this task is easily and effortlessly solved even by young children, it is challenging for machines since it is unclear which features constitute the "objectness" of an item. We present an approach that follows the mechanisms of the human visual system based on segmentation and saliency and show that our approach outperforms several state-of-the-art methods for object discovery. We show that the approach is applicable as well to RGB-D sequences and can be used to obtain 3D object models that are incrementally improved as soon as new information from other viewpoints is available.