Boosting 3D Foundation Models with Featureless Pose Optimization

We introduce Edge-based Pose Optimization (EPO),a trackless geometric optimization framework specifically designed toboost the Structure-from-Motion reconstructions generated by 3D Foun-dation Models. These models achieve rapid inference by bypassing thetime-consuming feature extraction and matching stages of traditionalpipelines, where explicit correspondences between each 3D point andmultiple images, referred to as tracks, are established. However, their ge-ometric accuracy currently falls short of traditional pipelines. While thiscan be addressed in a post-processing step via Bundle Adjustment-likerefinement, doing so requires extracting feature tracks, thus defeatingthe original speed advantage. Instead, our fully differentiable frameworkuses edge map alignment as a proxy for geometric optimization, avoid-ing feature extraction and track construction entirely. Through exten-sive evaluation across multiple datasets and tasks, we demonstrate thatEPO matches or outperforms Bundle Adjustment-like methods while re-quiring significantly lower runtime and memory. Notably, its reducedmemory footprint makes EPO suitable for consumer-grade hardware,where competing refinement methods cannot run. Code is available athttps://github.com/mattiadurso/EPO.