Solv3R: 다중 뷰 솔버로서의 3D 루프형 트랜스포머
기존의 single forward pass 방식이 아닌, 반복적으로 카메라 pose와 scene geometry를 정제하는 classical structure-from-motion의 장점을 살린 3D Looped Transformer인 Solv3R를 제안한다. 이 모델은 반복 횟수와 무관한 update operator를 사용하여, 이미 정제된 뷰와 새로 도입된 뷰를 통합적으로 업데이트할 수 있게 한다. 이를 통해 고정된 이미지 컬렉션의 정밀도를 높이는 refinement와 새로운 관측 데이터의 seamless한 integration을 모두 지원한다.
Classical structure-from-motion reconstruction proceeds iteratively, refining camera poses and scene geometry from a current estimate. This allows maps to easily expand without recomputing already computed views, or geometry estimates to improve as more observations are provided. Most geometric foundation models, however, predict entire geometry in a single forward pass or update a recurrent state as frames arrive sequentially, making this expansion and refinement non-trivial. We introduce Solv3R, a looped transformer that learns to continue solving multi-view geometry from its current state. Importantly, its update operator is independent of the iteration count, allowing previously refined and newly introduced views to be jointly updated regardless of how ``solved'' they are. We train the operator to preserve accurate reconstructions under continued iteration and correct inaccurate states, enabling refinement of fixed image collections and integration of new observations. Representati