Robust light fields denoising with S²N2N
Tal Kozakov, Omer Hazan, Adir Hazan, Adrian Stern · 2025
We have recently introduced the Single Shot Noise2Noise (S²N2N) framework for denoising Light Fields (LF) captured by Integral Imaging (InI), which is robust against different types and intensities of noise at any arbitrary exposure. S²N2N implicitly learns the noise type and intensity from the captured LF. In this paper, we further investigate S²N2N and introduce several improvements, including integration with a Visual Image Transformer (ViT). We test the method using both synthetic and real-world datasets, demonstrating significant improvements in denoising performance, particularly in high-noise environments. Our results reveal that this unsupervised denoising method has significant potential for real-world 3D imaging applications, offering robust performance without the need for explicit noise models.