Xinyi Sui

Date of Award

6-10-2025

Document Type

Thesis - SCU Access Only

Publisher

Santa Clara : Santa Clara University, 2025

Degree Name

Master of Science (MS)

Department

Computer Science and Engineering

First Advisor

Nam Ling

Abstract

Point cloud upsampling is a critical task in 3D vision, aiming to reconstruct dense and uniform point sets from sparse and irregular inputs. Existing interpolation-based methods offer high efficiency but suffer from limited reconstruction fidelity, while learningbased approaches improve accuracy at the cost of increased complexity. To address this tradeoff, we propose PU-ResInterp, a lightweight and interpretable framework for point cloud upsampling that integrates residual interpolation with a dual-branch architecture. The main branch employs a DGCNN-based encoder enhanced with CBAM attention modules to capture fine-grained geometric features, while the condition branch uses a PointNetstyle lightweight encoder to extract structural priors. These two branches are fused via FiLM modulation to enable semantic alignment. A two-stage decoder is then applied: a coarse decoder generates a preliminary upsampled point cloud, followed by a residual refinement module that predicts high-quality offsets to restore detailed geometry. Extensive experiments on the PU1K dataset demonstrate that PU-ResInterp achieves competitive performance with significantly fewer parameters and faster inference compared to existing methods. Our approach strikes a compelling balance between accuracy, interpretability, and computational efficiency, making it suitable for practical 3D applications.

Download

Share

COinS