Curvature-enhanced Neural Subdivision

Research output: Contribution to conferencePaperAcademic

81 Downloads (Pure)

Abstract

Subdivision is an important and widely used technique for obtaining dense meshes from coarse control (triangular) meshes for modelling and animation purposes. Most subdivision algorithms use engineered features (subdivision
rules). Recently, neural subdivision successfully applied machine learning to the subdivision of a triangular mesh. It uses a simple neural network to learn an optimal vertex positioning during a subdivision step. We propose an extension to the neural subdivision algorithm that introduces explicit curvature information
into the network. This makes a larger amount of relevant information accessible which allows the network to yield better results. We demonstrate that this modification yields significant improvement over the original algorithm, in terms of both Hausdorff distance and mean squared error.
Original languageEnglish
Publication statusPublished - 31-May-2023
EventThe 36th International Conference on Computer Animation and Social Agents - Limassol, Cyprus
Duration: 29-May-202331-May-2023

Conference

ConferenceThe 36th International Conference on Computer Animation and Social Agents
Abbreviated titleCASA 2023
Country/TerritoryCyprus
CityLimassol
Period29/05/202331/05/2023

Fingerprint

Dive into the research topics of 'Curvature-enhanced Neural Subdivision'. Together they form a unique fingerprint.

Cite this