LAF-NeRF: Learning Artifact-Free Neural Radiance Fields from Un-Curated Image Collections with Corruptions

The quality of the training images significantly influences the effectiveness of photorealistic view synthesis using Neural Radiance Fields (NeRF) techniques. However, casually captured image sets frequently exhibit a range of corruptions. These corruptions can result in artifacts like floaters or incorrect geometry when these images are utilized in NeRF processes. Current methodologies to address image corruption in NeRF are largely tailored to specific corruption types, constraining their applicability. In this paper, we propose a novel approach for learning artifact-free NeRFs in the presence of multiple types of corruption. Our method hinges on the observation that the volume density distribution along a ray can indicate whether the ray has been cast through a corrupted image region. Using a novel dataset containing a wide range of corruptions, we show that our proposed technique substantially enhances the quality and effectively reduces image artifacts. Our method improves the baseline models by close to 10 percentage points.