SplatDoF : Real-Time Depth of Field for Gaussian Splatting

Sammanfattning

Depth of Field (DoF) is an optical effect that is often used in various disciplines from photography over movies to computer graphics to draw the viewer’s attention to certain objects and give depth cues. While there is a wide range of techniques that are capable of artificially creating the effect in digital real-time applications, they face a significant limitation: Their reliance on depth information, which is often only available in a single layer, creates issues with transparencies. This work showcases a proof of concept of how 3D Gaussian Splatting could be used as an avenue to address these issues and create a physically-based DoF effect through exploitation of the radial nature of Gaussian Splats. Results of an evaluation with common image similarity metrics indicate that the concept is viable in principle, but to achieve full physical accuracy further research into the alpha blending of splats is required.