NoiseSDF2NoiseSDF: Learning Clean Neural Fields from Noisy Supervision
Researchers have developed a new method called NoiseSDF2NoiseSDF to improve the reconstruction of 3D neural fields from noisy point cloud data. This technique extends the Noise2Noise paradigm from 2D images to 3D, enabling the learning of clean surface estimations even when the input data is imperfect. By minimizing the mean squared error between noisy SDF representations, the network implicitly denoises and refines the surface, showing significant improvements on various benchmark datasets. AI
IMPACT Improves 3D reconstruction quality from imperfect data, potentially aiding applications in robotics and virtual reality.