| dc.contributor.advisor | Wörgötter, Florentin Prof. Dr. | |
| dc.contributor.author | Teich, Florian | |
| dc.date.accessioned | 2021-06-18T11:55:54Z | |
| dc.date.available | 2021-06-24T00:50:09Z | |
| dc.date.issued | 2021-06-18 | |
| dc.identifier.uri | http://hdl.handle.net/21.11130/00-1735-0000-0008-5867-2 | |
| dc.identifier.uri | http://dx.doi.org/10.53846/goediss-8667 | |
| dc.language.iso | eng | de |
| dc.rights.uri | http://creativecommons.org/licenses/by-nc-nd/4.0/ | |
| dc.subject.ddc | 510 | de |
| dc.title | Graph-based Object Understanding | de |
| dc.type | doctoralThesis | de |
| dc.contributor.referee | Wörgötter, Florentin Prof. Dr. | |
| dc.date.examination | 2021-06-01 | |
| dc.description.abstracteng | Computer Vision algorithms become increasingly prevalent in our everyday lives. Especially recognition systems are often employed to automatize certain tasks (i.e. quality control). In State-of-the-Art approaches global shape char acteristics are leveraged, discarding nuanced shape varieties in the individual parts of the object. Thus, these systems fall short on both learning and utilizing the inherent underlying part structures of objects. By recognizing common substructures between known and queried objects, part-based systems may identify objects more robustly in lieu of occlusion or redundant parts. As we observe these traits, there are theories that such part-based approaches are indeed present in humans. Leveraging abstracted representations of decomposed objects may additionally offer better generalization on less training data. Enabling computer systems to reason about objects on the basis of their parts is the focus of this dissertation. Any part-based method first requires a segmentation approach to assign object regions to individual parts. Therefore, a 2D multi-view segmentation approach for 3D mesh segmentation is extended. The approach uses the normal and depth information of the objects to reliably extract part boundary contours. This method significantly reduces training time of the segmentation model compared to other segmentation approaches while still providing good segmentation results on the test data. To explore the benefits of part-based systems, a symbolic object classification dataset is created that inherently adheres to underlying rules made of spatial relations between part entities. This abstract data is also transformed into 3D point clouds. This enables us to benchmark conventional 3D point cloud classification models against the newly developed model that utilizes ground truth symbol segmentations for the classification task. With the new model, improved classification performance can be observed. This offers empirical evidence that part segmentation may boost classification accuracy if the data obey part-based rules. Additionally, prediction results of the model on segmented 3D data are compared against a modified variant of the model that directly uses the underlying symbols. The perception gap, representing issues with extracting the symbols from the segmented point clouds, is quantified. Furthermore, a framework for 3D object classification on real world objects is developed. The designed pipeline automatically segments an object into its parts, creates the according part graph and predicts the object class based on the similarity to graphs in the training dataset. The advantage of subgraph similarity is utilized in a second experiment, where out-of-distribution samples ofobjects are created, which contain redundant parts. Whereas traditional classification methods working on the global shape may misinterpret extracted feature vectors, the model creates robust predictions. Lastly, the task of object repairment is considered, in which a single part of the given object is compromised by a certain manipulation. As human-made objects follow an underlying part structure, a system to exploit this part structure in order to mend the object is developed. Given the global 3D point cloud of a compromised object, the object is automatically segmented, the shape features are extracted from the individual part clouds and are fed into a Graph Neural Network that predicts a manipulation action for each part. In conclusion, the opportunities of part-graph based methods for object understanding to improve 3D classification and regression tasks are explored. These approaches may enhance robotic computer vision pipelines in the future. | de |
| dc.contributor.coReferee | May, Wolfgang Prof. Dr. | |
| dc.contributor.thirdReferee | Damm, Carsten Prof. Dr. | |
| dc.contributor.thirdReferee | Kurth, Winfried Prof. Dr. | |
| dc.contributor.thirdReferee | Waack, Stephan Prof. Dr. | |
| dc.contributor.thirdReferee | Yahyapour, Ramin Prof. Dr. | |
| dc.subject.eng | 3D | de |
| dc.subject.eng | Graphs | de |
| dc.subject.eng | Segmentation | de |
| dc.subject.eng | Classification | de |
| dc.identifier.urn | urn:nbn:de:gbv:7-21.11130/00-1735-0000-0008-5867-2-2 | |
| dc.affiliation.institute | Fakultät für Mathematik und Informatik | de |
| dc.subject.gokfull | Informatik (PPN619939052) | de |
| dc.description.embargoed | 2021-06-24 | |
| dc.identifier.ppn | 1760882224 | |