Deep Bundle Adjustment: 590x smaller than VGGT, initialization-free 3D reconstruction beats COLMAP beyond 15 views

Singularitarian · x · 2026-11-18

Daniel Cremers' TUM group (with Munich Center ML) present Deep Bundle Adjustment: accurate and fast 3D reconstruction without initialization. The model is 590x smaller than VGGT while scoring 97.4 vs 99.3 RRA@15 on CO3Dv2, and is faster than COLMAP/GLOMAP beyond 15 views.

Original post →

More from Research

Research channel →