Speaker: Martin Čavarga
Abstract:
This talk builds on the previously introduced Incremental Mesh Builder (IMB) framework for progressive Level-of-Detail (LOD) mesh reconstruction from large 3D datasets. The current focus is on evaluating the viability of a geometry-aware stochastic vertex sampling process as a foundation for adaptive mesh reconstruction — balancing simplification speed with feature preservation.
The latest developments test this sampling approach on both structured and noisy 3D datasets, using simplified artificial heightmaps and real-world scan-derived meshes. Performance results include comparisons between uniform and softmax-based sampling, along with LOD convergence times and visual fidelity metrics.
The talk will also outline the ongoing effort to shape this framework into a lightweight, open-source application tailored to experts in 3D scanning, offering an efficient and scalable way to browse massive raw mesh files without full preprocessing.
Bio:
Martin Čavarga specializes in combinatorial geometries in Euclidean space, particularly mesh surfaces and point cloud or voxel 3D data processing. His work includes CAD solid algorithms and geometric data structures.
His expertise lies in developing numerical simulation frameworks using FEM or FVM on discrete geometries, grounded in differential geometry principles. Current research focuses on robust multi-manifold shrink-wrapping for 2D and 3D data, and incremental simplification plugins for 3D applications. In his long-time industry experience, he also contributes to an automated polyhedral wall and finish element system for a civil engineering and architectural commercial application.