✦
Summary
GeoWeaver is a long-sequence 3D reconstruction framework for turning RGB video into globally consistent cameras and point clouds. It addresses the scale drift, pose errors, and chunk misalignment that appear when long sequences cannot be processed jointly.
Key Features
Long-sequence RGB 3D reconstruction.
Chunk-wise depth and camera prediction.
Hierarchical test-time adaptation.
Global Sim(3) chunk alignment.
Coarse-to-fine camera refinement.
Affine depth correction.
Dense correspondence constraints.
Globally consistent cameras and point clouds.
About GeoWeaver
A Geometric Prior Model predicts chunk-wise depth, confidence, and camera parameters, then hierarchical test-time adaptation performs sequential initialization, global Sim(3) alignment, and coarse-to-fine refinement. Dense correspondences and a robust CDF-style objective jointly optimize weighted 2D reprojection and 3D consistency residuals.
GeoWeaver is useful for indoor mapping, outdoor reconstruction, driving scenes, urban-scale capture, visual localization, and 3D asset creation from long videos. Its test-time adaptation remains compatible with different geometric prior models rather than being tied to one specific backbone.