Smooth Textured Surface Reconstruction from Point Cloud Rendered Images as well as Path Traced Images using Variational Methods

Abstract: In this work, we propose total variation-based methods for smoothing textured surfaces in point-based rendering and reducing noise in Monte Carlo-rendered images. Initially, we survey the challenges and existing state-of-the-art methodologies in these two research domains.

Subsequently, we delve into the details of our proposed total variational models, each aimed at smoothing point-rendered textured surfaces and reducing noise in Monte Carlo-rendered images, respectively. For smoothing textured surfaces in point-based rendering, our model incorporates geometric features and is then combined with an advanced Pull-Push method. This combined approach enables us to effectively fill gaps and smooth discontinuous surfaces. The models tailored for denoising Monte Carlo-rendered images leverage noise-free auxiliary features and noise estimation techniques. Our approach efficiently eliminates noise while preserving crucial image features. We conduct comprehensive comparison experiments against existing state-of-the-art techniques to evaluate the effectiveness of our methods. Although our implementations are currently offline, both the smoothing and denoising processes can be achieved within a few iterations. Given the simplicity of our approach’s implementation, we foresee the potential for a GPU-based implementation, paving the way towards real-time applications.