Authors:
Jannik Koch
1
;
2
;
Laura Haraké
1
;
Alisa Jung
2
and
Carsten Dachsbacher
2
Affiliations:
1
Fraunhofer Institute of Optronics, System Technologies and Image Exploitation IOSB, Ettlingen, Germany
;
2
Karlsruhe Institute of Technology, Karlsruhe, Germany
Keyword(s):
Generative Models, Shape Synthesis, Graph Neural Networks, Physical Constraints, Measure of Infeasibility.
Abstract:
StructureNet is a recently introduced n-ary graph network that generates 3D structures with awareness of geometric part relationships and promotes reasonable interactions between shape parts. However, depending on the inferred latent space, the generated objects may lack physical feasibility, since parts might be detached or not arranged in a load-bearing manner. We extend StructureNet’s training method to optimize the physical feasibility of these shapes by adapting its loss function to measure the structural intactness. Two new changes are hereby introduced and applied on disjunctive shape parts: First, for the physical feasibility of linked parts, forces acting between them are determined. Considering static equilibrium, compression and friction, they are assembled in a constraint system as the Measure of Infeasibility. The required interfaces between these parts are identified using Constructive Solid Geometry. Secondly, we define a novel metric called Hover Penalty that detects
and penalizes unconnected shape parts to improve the overall feasibility. The extended StructureNet is trained on PartNet’s chair data set, using a bounding box representation for the geometry. We demonstrate first results that indicate a significant reduction of hovering shape parts and a promising correction of shapes that would be physically infeasible.
(More)