Gaussian
Splatting

I specialise in high-fidelity 3D scene reconstruction, focusing on surface estimation and mesh extraction using Gaussian Splatting. Gaussian Splatting represents a scene as a collection of Gaussian primitives and renders it in real time via a rasterisation-based pipeline, rather than requiring neural rendering. These Gaussians, being differentiable volumetric representations, enable compact optimisation, resulting in efficient training times whilst providing exceptional visual quality.

Extracting high-quality surface meshes from Gaussian Splatting remains a challenge due to the loose geometric alignment of ellipsoidal Gaussians. 2DGS proposes the use of 2D Gaussian primitives (disks) as surface elements, which significantly improves quality. Our modified representation using flattened 2D Gaussians improves surface alignment and geometric accuracy. This leads to high-fidelity, textured mesh reconstruction from a limited set of images, capable of capturing facial details, hair and mouth interior effectively.

This work was presented at ACM SIGGRAPH CVMP 2024 in London, UK and at the AniNex Workshop, CASA 2025 in Strasbourg, France. Please see the publications page.

In collaboration with Visualskies Ltd, I conducted a series of experiments with progressively fewer input images (from 116 down to just 9) to evaluate the technique's robustness and ability to maintain fidelity under sparse views. Notice the level of detail retained, even when trained with only 9 images.

The technique also extends to outdoor scenes. This model is trained on a subset of "tree stump" (77 images, 2× downsampled) from the Mip-NeRF 360 dataset. The final splat is only 5 MB, smaller than a single image.