CVPR Daily - Tuesday

This paper turns the best-case scenario on its head , aiming to address the challenge of novel-view synthesis based on a neural field representation using as few as two to three views and noisy poses. The potential applications of this technology are significant. In robotics , it could capture 3D reconstructions from a few images, saving time and resources. It could also be used in AR or VR applications , such as remodeling apartments, with users feeding images of furniture or other objects into the model to generate novel- view renderings and visualize how it would look in their living room. After completing a literature review, Prune discovered that several works had already tackled sparse input views and found solutions to help with the overfitting problem. However, these works typically used perfect ground truth poses. There had also been works on a joint pose-NeRF refinement, but they assumed dense images. 4 DAILY CVPR Tuesday Highlight Presentation