Real-time Simultaneous Localization and Photorealistic Mapping for Monocular, Stereo, and RGB-D Cameras

Photo-SLAM is a novel SLAM framework that maintains a hyper primitives map to efficiently optimize tracking using a factor graph solver and learn the corresponding mapping by backpropagating the loss between the original images and rendering images. It introduces geometry-based densification and Gaussian-Pyramid-based learning to enhance online photorealistic mapping performance.