Robust 3D Gaussian Splatting for Images with Accidental Shadows

In this paper, we propose a method for optimizing 3D Gaussians that is robust against artifacts caused by accidental shadows in images. For images without accidental shadows, the standard 3D Gaussian splatting method achieves very high-quality results. However, when accidental shadows are present, they introduce artifacts that degrade model quality. Excluding accidental shadow regions from the learning process can greatly reduce these artifacts, but accurately detecting shadow regions is challenging. This task becomes even more complex when distinguish accidental shadows from ordinary ones in cases where camera position or orientation changes, as in images used for the 3D Gaussian splatting. To address these challenges, we introduce a method that identifies and excludes accidental shadow regions during the 3D Gaussian splatting learning process. Our approach utilizes ground truth images alongside rendered images to detect accidental shadow regions for generating shadow region masks, which are then applied to effectively mitigate artifact formation. Experimental results demonstrate that our method performs well on both shadow-free and shadow-affected images, maintaining the optimization speed characteristic of 3D Gaussian splatting.