Temporal 3D scene representations from sparse data

This proposal aims to explore novel techniques deriving from Neural Radiance Fields (NeRFs) and 3D Gaussian Splatting in producing time-dependent 3D explorable environments from sparse imagery. This is a very active field of research in which methods are being explored for either representing environments that change gradually over time, or generating a representation from limited source images.

However, there is a real need for capturing environments with sudden changes in geometry in the context of having limited source images with various lighting conditions, which this proposal aims to investigate.

The plan would be to start by conducting a literature review of the latest techniques in NeRFs and Gaussian Splatting. Then a controlled test environment, with variable lighting and geometry, would be set up to test combinations of methods and proposed alternatives with the aim of developing a pipeline with which to process datasets and produce time-dependent radiance fields.

This pipeline could then be validated and refined against real-world datasets. A particular type of dataset useful for this project would be images of a glacier across months and years, as it would contain the relevant challenges being addressed.

The benefits of the project are two-fold. One is the investigation of new techniques in capturing and representing changing environments, to contribute new insights to the field. The other major benefit is to glaciologists by providing a method by which they can monitor the historical and future recession of glaciers, and with which they can better understand the effects of climate change.