Point clouds from LiDAR, depth cameras, and reconstruction pipelines are inherently unordered and irregular. Standard convolutions assume a grid; geometric deep learning must instead define derivatives and neighbourhoods that respect surface and spatial structure.
In collaboration with colleagues at TU Delft, I contributed to DeltaConv, which introduces anisotropic operators inspired by discrete exterior calculus so that learned filters can align with local geometry rather than treating all directions equally. The line of work spans an arXiv preprint, the ACM TOG journal version, and open-source implementations for reproducibility.
Anisotropic operators & CNN architectures for point clouds
We design convolution-like layers whose stencils adapt to estimated tangent and normal structure, improving feature extraction for segmentation, classification, and scene understanding. The goal is to make 3D networks more data-efficient and faithful to curvature while remaining compatible with modern training stacks.
Related publications
3D mapping, recognition, and applications
Better point-cloud networks support autonomous systems, digital twins, and robotics. While my published focus is on the operator design itself, the motivation includes robust object recognition and real-time understanding of large scans where anisotropy and scale vary across the scene.
Related publications
All papers in this area
DeltaConv line of work; newest first.
DeltaConv: Anisotropic Geometric Deep Learning with Exterior Calculus
Additional Scholar listing (related to DeltaConv / TOG 2022), 2022