3D model based object recognition using an assembly of discrete primitives

Abstract

Recognizing different categories of objects is crucial for human interaction with the physical environment. This goal is as important to any technology that needs to interact autonomously within a similar setting. Current object categorization systems have made significant advances in performance in the past years, especially with the onset of visual based methods. However such systems can be challenged when it comes to categories of objects whose instances vary significantly in geometry, which inherently leads to more complex changes in object appearance. On the other hand methods that use a 3D model representation can better handle complex geometries and are being revisited in an increasing number of studies. 3D model based approaches however have been mostly limited to exemplar based methods that look for specific geometric similarities and do not capture a very generic model. This can prove to be a limiting factor when it comes to instances that exhibit large intra-class variance. We hypothesize that segmenting 3D objects into parts and investigating the part-to-part relationships can reveal a powerful cue that is useful for object categorization. This hypothesis is based on well established works demonstrating that object category is linked to function, and that function is linked to the overall geometric strncture. In this work we associate the geometric properties directly to category by investigating object structure rather than individual part properties. To exploit these cues we present a framework which starts by segmenting category labeled 3D models into primitive parts. Part-to-part relationships are then derived and used to learn a category model. It has been shown that for 2D object categorization, part-to-part relationships provide better categorization perfonnance than individual part properties; in this work we investigate this concept with respect to 3D parts. Using graph matching techniques, 3D model instances are aligned and used to build compact category models without requiring any prior part labeling. We demonstrate that our framework is able to automatically capture and learn 3D part-to-paii relationships and perform category recognition. We test the implemented algorithm using 745 different 3D models which are obtained from the Toyohashi shape benchmark and are manually categorized into 6 different object categories. Results demonstrate that the majority of the classified instances were associated to the correct category.