Neural radiance field methods, including approaches such as 3D Gaussian Splatting, have significantly advanced the state of the art in novel view synthesis and 3D scene representation. They enable highly realistic rendering from sparse observations, but this quality comes at the cost of substantial memory usage and computational complexity. These limitations become particularly critical in scenarios that require real-time performance or deployment on specialized visualization systems such as light-field displays.

Research Objectives: This thesis focuses on developing efficient compression strategies for radiance field representations to reduce memory requirements while preserving visual fidelity. Rather than treating compression as a purely technical problem, the work will explicitly consider the trade-offs between compression rate and perceived quality. Evaluation will play a key role throughout the project. In addition to standard objective metrics such as PSNR, SSIM, and perceptual measures, the thesis will place particular emphasis on perceptual quality assessment in a 3D context. Using a light-field display system available for experiments, the student will investigate how compression artifacts affect depth perception, visual comfort, and overall realism. This aspect is especially important, as artifacts that appear minor in 2D projections can become significantly more noticeable in 3D visualization.


Methodology and Expected Contributions: The research will involve designing and evaluating compression techniques for radiance field representations, with experiments conducted across different datasets and compression settings. Both objective metrics, such as PSNR and SSIM, and perceptual measures will be used for evaluation. Subjective experiments will be conducted to assess the impact of compression artifacts in realistic viewing conditions.
This thesis is expected to contribute to the development of memory-efficient radiance field representations and to deepen understanding of the relationships among compression, perceptual quality (Rate-Distortion tradeoff), and visualization conditions. The outcomes will support the design of compression methods that are better aligned with human perception, particularly in the context of advanced 3D display systems.

Contact Saeed.Mahmoudpour@vub.be for more information.

Kerbl, Bernhard, et al. ŗd gaussian splatting for real-time radiance field rendering." ACM Trans. Graph. 42.4 (2023): 139-1.
Zhang, Richard, et al. "The unreasonable effectiveness of deep features as a perceptual metric." Proceedings of the IEEE conference on computer vision and pattern recognition. 2018.
Bagdasarian, Milena T., et al. ŗdgs. zip: A survey on 3d gaussian splatting compression methods." Computer Graphics Forum. Vol. 44. No. 2. 2025.
Chen, Yihang, et al. "Hac: Hash-grid assisted context for 3d gaussian splatting compression." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024.

Strong programming skills (preferably Python, and familiarity with PyTorch or similar frameworks)
Basic knowledge of computer vision and/or machine learning
Familiarity with 3D graphics or rendering is a plus, but not strictly required
Interest in experimental research and evaluation (including perceptual assessment)