Object Understanding Through Abstraction Inspired by Compositionality and Connectionism – From Noisy Scene Point Clouds to Objects and Shape Reasoning
- This thesis focuses on object perception. Particularly we focus on the description and representation of object shape information for detection and reasoning purposes by considering spatial information in form of RGBD data. Such RGBD data can be expressed by point clouds in which points are independently organized, but a subset or as a whole, may contain relevant information on different semantic levels and granularity. Inspired
by visuoperceptual principles such as compositionality and connectionism, i.e. complex structures can be expressed by the relation and interplay of simpler ones, the presented work proposes a hierarchical and data-driven abstraction process to reveal visual patterns and persistent structures: from generating building blocks of our surroundings over detecting potential object candidates to reasoning about the semantics of object shapes. In the course of our work, we aim on generic approaches such as detection of unknown objects in unstructured environments or the classification of object shape types instead of the recognition of individual object instances. Our eventual goal is a data-driven unsupervised conceptualization of shape commonalities regarding object appearance. Therein we aim to reduce supervision in form of handcrafted labeled data and to learn machine-centric prediction models that are solely based on the given data without incorporating knowledge that is not inferable from given data or is biased by supervision.
Further on, we specifically aim on the applicability in real-world scenarios coping with challenging conditions as in form of noisy sensor data to partial observations and occlusions. The presented work showed its applicability in autonomous unloading of goods from shipping containers or in shelf replenishment in retail scenarios. Additionally, our work was applied to a tool substitution framework.