Haniye Merrikhi, Hossein Ebrahimnezhad,
Volume 20, Issue 4 (December (Special Issue on ADLEEE) 2024)
Abstract
Robots have become integral to modern society, taking over both complex and routine human tasks. Recent advancements in depth camera technology have propelled computer vision-based robotics into a prominent field of research. Many robotic tasks—such as picking up, carrying, and utilizing tools or objects—begin with an initial grasping step. Vision-based grasping requires the precise identification of grasp locations on objects, making the segmentation of objects into meaningful components a crucial stage in robotic grasping. In this paper, we present a system designed to detect the graspable parts of objects for a specific task. Recognizing that everyday household items are typically grasped at certain sections for carrying, we created a database of these objects and their corresponding graspable parts. Building on the success of the Dynamic Graph CNN (DGCNN) network in segmenting object components, we enhanced this network to detect the graspable areas of objects. The enhanced network was trained on the compiled database, and the visual results, along with the obtained Intersection over Union (IoU) metrics, demonstrate its success in detecting graspable regions. It achieved a grand mean IoU (gmIoU) of 92.57% across all classes, outperforming established networks such as PointNet++ in part segmentation for this dataset. Furthermore, statistical analysis using analysis of variance (ANOVA) and T-test validates the superiority of our method.
Amirreza Amirfathiyan, Hossein Ebrahimnezhad,
Volume 20, Issue 4 (December (Special Issue on ADLEEE) 2024)
Abstract
This paper presents an application of deep learning in computer graphics, utilizing learn-based networks for 3D shape matching. We propose an efficient method for shape matching between 3D models with non-isometric deformation. Our method organizes intrinsic and directional attributes in a structured manner. For this purpose, we use a hybrid feature derived from Diffusion-Net and spectral features. In fact, we combine learned-based intrinsic properties with orientation-preserving features and demonstrate the effectiveness of our method. We achieve this by first extracting features from Diffusion-Net. Then, we compute two maps based on the functional map networks to obtain intrinsic and directional features. Finally, we combine them to achieve a desired map that can resolve symmetry ambiguities on models with high deformation. Quantitative results on the TOSCA dataset indicate that the proposed method achieves lowest average geodetic error of 0.0023, outperforming state-of-the-art methods and reducing the error by 70.66%. We demonstrate that our method outperforms similar approaches by leveraging an accurate feature extractor and effective geometric regularizers, allowing for better handling of non-isometric shapes and resulting in reduced matching errors.
