acezero

Why we included this project

Most structure-from-motion tools recover camera poses by matching features across images. ACE0 takes a different route: it learns a multi-view consistent, implicit scene representation and estimates the camera parameters for a set of images from that, sidestepping the feature-matching pipeline entirely. The codebase is useful beyond the ECCV 2024 paper. You can refine existing poses, start from a partial reconstruction, feed it RGB-D input, or use the recovered poses to train NeRF and Gaussian splat models. It also bundles the standard ACE (CVPR 2023) workflow and optional reconstruction priors from the ICCV 2025 follow-up, so comparing classic and learning-based approaches is straightforward. Teams that need dependable camera poses for AR, robotics, or photogrammetry work will find the incremental learning approach and the included benchmark scripts directly applicable.