Gaussian splatting has recently been used for realistic representations of 3D scenes for real-time rendering. However, this representation does not natively allow rendering of a reconstructed object under novel lighting. This thesis will develop a novel multi-dimensional gaussian splatting representation that will contain scene information for many sampled light directions. Such a representation can be created either using a specialized scanning device (e.g. light stage) or by pre-rendering digital scene with many illumination directions and then reconstructing this scene into gaussian splatting representation. Then, the rendering under novel illumination can be calculated as a liner combination of multiple reconstructed gaussian splatting data points. The spatial locations of points in the point cloud should be the same for each light direction and thus the gaussian splatting representation will be optimized jointly for all light directions. The implementation will be done in python and OpenGL.  The visual quality of the rendered image under novel lighting will be evaluated quantitatively by calculating difference to ground truth solution.

Type: Master thesis

For more information please contact Peter Kán