3D Deep Learning for Geometric Analysis and Understanding

← Back to Home

My research in 3D deep learning, initiated in 2020, was motivated by the remarkable success of deep learning in 2D vision tasks such as image recognition, segmentation, and generative modeling. While these advances demonstrated the power of data-driven representation learning, extending similar success to 3D domains presents fundamental challenges. Unlike images defined on regular grids with a global coordinate system, 3D data are typically unstructured and irregular, including point clouds, meshes, and implicit fields. They lack canonical parameterizations, consistent sampling patterns, and simple convolutional structures. To address this gap between 2D and 3D learning, my work integrates classical digital geometry processing and discrete differential geometry with modern deep learning techniques for geometric understanding and generation. Building upon these geometric foundations, we develop neural models for point cloud denoising, completion, semantic segmentation, and single- or multi-view reconstruction. A central theme of this research is embedding geometric structure into learning frameworks, enabling robustness, scalability, and faithful surface representation.

Neural Implicit Representations

Across my recent work on neural implicit representations, my goal has been to turn learned continuous fields, especially distance-based functions, into dependable geometric primitives: expressive enough to capture fine-scale real-world detail, and reliable enough to support downstream geometry processing. A major focus is learning unsigned distance fields (UDFs) directly from raw point clouds for high-fidelity reconstruction. GeoUDF (ICCV 2023) introduces geometry-guided learning to stabilize UDF prediction from noisy and incomplete scans; DEUDF (AAAI 2025) targets the persistent challenge of detail preservation; and LoSF-UDF (CVPR 2025) leverages local shape functions to reduce sensitivity to the training distribution and improve generalization. To bridge the gap from “a neural field” to usable geometry, especially when sign information is unavailable, we also develop principled discretization and extraction methods, including DCUDF (SIGGRAPH Asia 2023) and DCUDF2 (TVCG 2025), which improve the stability, efficiency, and accuracy of zero level-set recovery from learned UDFs. We then push these representations to more challenging regimes (e.g., non-manifold and multi-material interfaces in MIND, NeurIPS 2025) and extend neural fields into tools for shape computation and understanding: NeuroGF (NeurIPS 2023) enables fast geodesic distance and path queries, while Q-MDF (TOG 2026) exploits distance-field structure to robustly approximate and discretize neural medial axes. Most recently, SharpNet (arXiv 2026) advances neural representation by introducing controlled non-differentiability into MLPs, enabling faithful modeling of sharp features and piecewise-smooth behavior. Together, these works form a coherent agenda of geometry-aware neural fields with controllable regularity and reliable extraction, making implicit representations practical foundations for reconstruction, analysis, and shape reasoning.

Point Cloud Denoising

This line of work revisits point cloud denoising, a classical geometry processing problem, through the lens of learning, while keeping geometric structure and interpretability at the core. Starting from feature-preserving displacement regression (Pointfilter, TVCG 2021), we progressed to trained iterative refinement (IterativePFN, CVPR 2023), joint point/normal reasoning (PCDNF, TVCG 2024), and adaptive stopping to avoid over- and under-denoising (ASDN, AAAI 2025), with extensions to scene-scale inputs (3DMambaIPF, AAAI 2025). We further explored geometry-grounded and generative formulations, including implicit-field guidance (TVCG 2025), invertible latent-space denoising (CVPR 2024), adaptive latent alignment for unseen real noise (LaPDA, TVCG 2026), and deterministic residual diffusion guided by geometric displacements (TVCG 2026), alongside learned structural priors (AAAI 2026). We also summarized the broader landscape of deep learning-based denoising in a survey (arXiv 2025).

Parameterization and Sampling

Topology Optimization and 3D Generation

Motion, Gesture, and Correspondence

Single- or Multi-view Reconstruction

Completion, Segmentation, and Understanding

Copyright Notice: These materials are provided for academic dissemination only. Copyright remains with the authors or respective copyright holders. Reproduction or redistribution may require prior permission.