Motivation
Surface reconstruction from noisy points has a intensely researched since it is an important problem in representing real-world geometry well as 3D models. Fitting 3D Gaussian Splats (3DGS) [1] to scanned RGB images results artifacts such as splats that are not well aligned to the surface. Existing methods try to learn better alignments, over-smoothing the resulting surface in the process.
Description
3DGS add direction and magnitude in all three dimensions over a simple point, aside from color and alpha. Classic Poisson Surface Reconstruction creates a signed distance function (SDF) with points and normals but cannot handle noise well. By integrating these noisy splats from 3DGS into a probabilistic SDF, we can use a method such as Stochastic Poisson Reconstruction [2] to aim for a more precice surface reconstruction.
[1] https://hal.science/hal-04088161/
[2] https://www.silviasellan.com/pdf/papers/stochastic-psr.pdf
Tasks (depending on PR/BA/DA and number of students)
- Define the probabilistic SDF from 3DGS data using the covariance matrices and alpha values similar to the Gaussian process here as mean and variance [2]
- Implement the computation of this probabilistic SDF
- Evaluate the quality of the surface reconstruction and compare the results with the baseline [2]
Requirements
- Knowledge of English language (source code comments and final report have to be in English)
- Knowledge of C++ is necessary, experience in statistics and CUDA or parallel programming a plus
Environment
The project should be implemented platform independent (Linux, Windows).
A bonus of €500/€1000 if completed to satisfaction within an agreed time-frame of 6/12 months (PR/BA or DA)