Los puntos clave no están disponibles para este artículo en este momento.
Recently, 3D Gaussian Splatting (3D-GS) has gained popularity in novel-view scene synthesis, achieving high fidelity and efficiency. However, the rendering quality of 3D-GS is still not up to the quality of the best methods in NeRF, and serious artifacts will occur when changing the sampling rate and changing the focal length or camera distance. Another significant problem is that interactive real-time rendering of complex scenes also has great limitations. To address this challenge, we have come up with a new rendering system, UE-GS, that blends 3D-GS with traditional rendering techniques in Unreal Engine 5 (UE5) to achieve higher quality visuals. Our approach draws inspiration from UE5's Level of Detail (LOD) technology. When changing the focal length or camera distance, we dynamically render LOD for different viewing levels, and replace the dilatation filter in 3D-GS with 2D EWA filtering, effectively alleviating the problems of aliasing and dilatation artifacts. Use octree to optimize spatial point data in complex scenes (40 million Gaussian +) and customize rendering components so that they are not limited by particle counts in UE5. We validated our approach on a variety of data sets and scenes that retain visual quality, achieving better rendering quality and faster training and reasoning speed than 3D-GS, while retaining the benefits of real-time performance. In addition, rendering in UE5 helps with scene editing in subsequent stages.
Jin et al. (Fri,) studied this question.