Document Type

Thesis

Degree Name

Master of Applied Computing

Department

Physics and Computer Science

Program Name/Specialization

Applied Computing

Faculty/School

Faculty of Science

First Advisor

Emad Mohammed

Advisor Role

Supervisor

Abstract

This thesis studies efficiency and stability challenges in Gaussian-splatting-based reconstruction under sparse supervision. In few-shot novel view synthesis, standard 3D Gaussian Splatting (3DGS) can overfit the limited training views and grow an unnecessarily large number of primitives due to limitations in its Adaptive Density Control (ADC) mechanism. This thesis introduces an error-driven reformulation of ADC that triggers densification using opacity gradients as a lightweight proxy for rendering error, and shows that such aggressive densification must be paired with delayed and conservative pruning to prevent destructive create--destroy cycles. When combined with depth-based geometric regularization, the resulting framework produces substantially more compact representations while maintaining competitive reconstruction quality on standard benchmarks.

Building on the same core idea of allocating capacity where the loss indicates high error sensitivity, this thesis also develops a sparse-view tomographic reconstruction method using radiometrically-correct Gaussian splatting for X-ray image formation (including integration-bias rectification) with density-gradient-driven density control designed to focus primitives on projection-domain error under sparse-view ambiguity. On the R2-Gaussian benchmark (15 synthetic volumes and three real-world objects evaluated at 25/50/75 views), the tomography method reduces primitive count and improves efficiency: on real data at 50 views and 30k iterations, the mean Gaussian count is reduced by up to 42% while maintaining competitive PSNR/SSIM, with representative runs showing 30--43% lower runtime and 33--43% lower peak allocated VRAM.

Convocation Year

2026

Convocation Season

Spring

Download

Share

COinS