PGI-Net: Deep geometry image representation for 3D point clouds via progressive parameterization and optimization

Cen, Jiaxin, Yue, Guanghui, Zhou, Wei, Mao, Xudong, Lin, Shujin, Su, Zhuo and Zhao, Baoquan 2026. PGI-Net: Deep geometry image representation for 3D point clouds via progressive parameterization and optimization. Presented at: MM '26: The 34th ACM International Conference on Multimedia, Rio de Janeiro, Brazil, 10 - 14 November 2026. MM '26: Proceedings of the 34th ACM International Conference on Multimedia. ACM International Conference on Multimedia (ACM MM). New York: ACM, pp. 234-243. 10.1145/3767308.3836020

PDF - Published Version Available under License Creative Commons Attribution. Download (13MB)

Abstract

The inherent irregularity and lack of structure in point cloud data present fundamental challenges for efficient processing, representation learning, and spatiotemporal modeling. Point Geometry Image (PGI) methods offer a promising direction by regularizing point clouds into structured 2D representations, yet existing approaches typically learn direct mappings from 3D point clouds to 2D grids, placing a heavy burden on the network to simultaneously discover global organization and local geometric refinement. This often results in unstable geometric correspondences and limited temporal consistency. We propose PGI-Net, a geometry-guided structured representation framework that converts 3D point clouds into compact and regular 2D Point Geometry Images through spherical and octahedral geometric initialization. A multi-stage progressive refinement strategy with adaptive resampling further improves local neighborhood preservation, global geometric fidelity, and temporal coherence across frames. The resulting representation is compact, structured, and reconstruction-consistent, while being readily compatible with standard 2D backbones for downstream analysis. Extensive experiments demonstrate that PGI-Net achieves strong geometric fidelity, structural consistency, and computational efficiency, while effectively supporting downstream tasks, including action recognition, temporal interpolation, and point cloud compression, establishing geometry-guided PGI construction as an effective paradigm for structured point cloud representation.

Item Type:	Conference or Workshop Item - published (Paper)
Date Type:	Published Online
Status:	Published
Schools:	Schools > Computational & Mathematical Sciences Schools > Computer Science & Informatics
Publisher:	ACM
ISBN:	9798400722134
Date of First Compliant Deposit:	13 November 2026
Last Modified:	13 Nov 2026 11:45
URI:	https://orca.cardiff.ac.uk/id/eprint/190899

Actions (repository staff only)

Edit Item