Learning-based Point Cloud Geometry Compression with Transformers and Scalable Coding

Luo, Zhe

Learning-based Point Cloud Geometry Compression with Transformers and Scalable Coding

Luo, Zhe

Permalink

Publication Type:: Thesis
Issue Date:: 2026

Open Access

Copyright Clearance Process

Recently Added
In Progress
Open Access

This item is open access.

Adobe PDF

Download thesisAdobe PDF (4.3 MB)

View statistics

Full metadata record

Field	Value	Language
dc.contributor.author	Luo, Zhe
dc.date.accessioned	2026-08-04T02:29:04Z
dc.date.available	2026-08-04T02:29:04Z
dc.date.issued	2026
dc.identifier.uri	http://hdl.handle.net/10453/195909
dc.description	University of Technology Sydney. Faculty of Engineering and Information Technology.	en_US.UTF-8
dc.description.abstract	Point clouds, as unstructured collections of 3D points, have become essential for representing complex geometries in applications ranging from autonomous driving to immersive media and cultural heritage preservation. However, their large data and irregular structures pose significant challenges in storage and transmission, motivating research into advanced compression techniques that balance rate-distortion (RD) performance while supporting scalable decoding and compressed-domain analysis. This thesis addresses these challenges through proposed learning-based point cloud compression (PCC) frameworks. Firstly, this thesis proposes a transformerbased autoencoder with local neighbor aggregation and global attention mechanisms to preserve intricate local and global features, achieving superior geometry coding and BD-Rate performance over recent baseline methods on ShapeNetCorev2. Secondly, this thesis introduces an edge-preserving compressed-domain classification approach using local attention to extract features directly from bitstreams, enhancing classification accuracy on ModelNet40 at reduced bitrates. Third, this thesis develops ScalaDAT, a density-aware scalable PCC method employing a tail-drop strategy for incremental decoding in a single training iteration, enhancing RD efficiency on widely-used datasets like SemanticKITTI and ShapeNet, outperforming benchmarks. This thesis’s contributions advance PCC by integrating and enhancing transformer architectures, attention-driven feature preservation, and scalable mechanisms, enabling scalable, efficient 3D data handling for real-life applications. These advancements not only mitigate computational and bandwidth constraints but also pave the way for seamless integration with downstream tasks, fostering innovations in 3D spatial data processing.	en_US.UTF-8
dc.format	Thesis (PhD)
dc.language.iso	en_US	en_US.UTF-8
dc.relation	https://opus.lib.uts.edu.au/bitstream/10453/195909/1/thesis.pdf
dc.rights	info:eu-repo/semantics/openAccess
dc.rights	The author owns the copyright in this thesis including all reproduction and reuse rights for the work. The work may not be altered without the permission of the copyright owner. Attribution is essential when quoting or paraphrasing from this thesis.
dc.rights	© 2026 Zhe Luo
dc.rights	au.edu.uts.lib/cph
dc.title	Learning-based Point Cloud Geometry Compression with Transformers and Scalable Coding	en_US.UTF-8
dc.type	Thesis
utslib.copyright.status	open_access	*

Abstract:

Point clouds, as unstructured collections of 3D points, have become essential for representing complex geometries in applications ranging from autonomous driving to immersive media and cultural heritage preservation. However, their large data and irregular structures pose significant challenges in storage and transmission, motivating research into advanced compression techniques that balance rate-distortion (RD) performance while supporting scalable decoding and compressed-domain analysis. This thesis addresses these challenges through proposed learning-based point cloud compression (PCC) frameworks. Firstly, this thesis proposes a transformerbased autoencoder with local neighbor aggregation and global attention mechanisms to preserve intricate local and global features, achieving superior geometry coding and BD-Rate performance over recent baseline methods on ShapeNetCorev2. Secondly, this thesis introduces an edge-preserving compressed-domain classification approach using local attention to extract features directly from bitstreams, enhancing classification accuracy on ModelNet40 at reduced bitrates. Third, this thesis develops ScalaDAT, a density-aware scalable PCC method employing a tail-drop strategy for incremental decoding in a single training iteration, enhancing RD efficiency on widely-used datasets like SemanticKITTI and ShapeNet, outperforming benchmarks. This thesis’s contributions advance PCC by integrating and enhancing transformer architectures, attention-driven feature preservation, and scalable mechanisms, enabling scalable, efficient 3D data handling for real-life applications. These advancements not only mitigate computational and bandwidth constraints but also pave the way for seamless integration with downstream tasks, fostering innovations in 3D spatial data processing.

Please use this identifier to cite or link to this item:

http://hdl.handle.net/10453/195909