Advancing Compression-Driven Approaches in 2D and 3D Vision

PhD Thesis Proposal Defence


Title: "Advancing Compression-Driven Approaches in 2D and 3D Vision"

by

Mr. Ka Leong CHENG


Abstract:

The exponential growth of visual data, including 2D images and 3D modalities 
such as videos and multi-view images, has placed a premium on efficient 
representation and processing techniques that extend beyond traditional data 
compression. This thesis explores compression-driven approaches as 
innovative solutions for 2D and 3D vision, leveraging compression not merely 
for data reduction but as a transformative tool to enhance efficiency, 
versatility, and quality across various vision tasks. By leveraging the 
principles of compression, this thesis addresses a range of challenges, 
including reversible image transformations, learned lossy compression, joint 
task optimization with compression, as well as efficient 3D editing with 
compact yet expressive representations.

To this end, we first develop frameworks that utilize invertible neural 
networks to encode and recover visual information with high fidelity, 
enabling tasks such as reversible image conversion and lossy image 
compression. By exploring the inherent invertibility of neural networks, we 
demonstrate that compression can serve as a reversible conduit for hiding 
and reconstructing multiple images within a single embedding, as well as 
improving the quality and efficiency of learned image compression codecs.

We then extend image compression methods beyond data reduction by 
investigating their synergy with other tasks. Specifically, we propose a 
joint learning framework for image compression and denoising, leveraging the 
innate denoising capabilities of compression models. This approach achieves 
superior results in both domains while revealing the latent connections 
between data reduction and noise suppression.

Finally, we expand the scope of compression to 3D vision, introducing a 
novel editing-friendly framework that encapsulates the appearance of 3D 
scenes into compact 2D canonical images. By treating the canonical image as 
a compressive representation of the 3D scene, this approach enables 
efficient 3D editing through standard 2D tools, eliminating the need for 
costly re-optimization while maintaining fidelity to the original scene.


Date:                   Wednesday, 8 January 2025

Time:                   2:00pm - 4:00pm

Venue:                  Room 5501
                        Lifts 25/26

Committee Members:      Dr. Qifeng Chen (Supervisor)
                        Prof. Raymond Wong (Chairperson)
                        Dr. Junxian He
                        Prof. Pedro Sander