The Splatter Image is an ultra-fast method for single- and few-view 3D reconstruction developed by the Visual Geometry Group at the University of Oxford. The method utilizes Gaussian Splatting, which has proven to be efficient in real-time rendering and fast training for multi-view reconstruction. In a monocular reconstruction setting, the Splatter Image uses a learning-based approach that only requires the feed-forward evaluation of a neural network for reconstruction. The innovative design of the Splatter Image involves mapping the input image to one 3D Gaussian per pixel, resulting in a Gaussian mixture that can be rendered quickly. The method also incorporates cross-attention views for incorporating multiple images during reconstruction. It achieves faster reconstruction and better results compared to more expensive baselines in terms of various metrics. The Splatter Image’s key challenge lies in designing a network that can accurately represent all sides of an object using 3D Gaussians. Surprisingly, the method achieves high-quality 360-degree reconstructions even when compared to slower methods. It uses a 2D image as a container for the 3D Gaussians, utilizing each pixel to store the parameters of one Gaussian. The Splatter Image successfully represents unobserved object parts and predicts occluded elements by allocating background
https://szymanowiczs.github.io/splatter-image