DeepBA: a 2M-parameter GNN learns bundle adjustment and beats prior learned SfM from a cold start
Singularitarian · x · 2026-11-18
A new preprint from TUM researchers presents Deep Bundle Adjustment (DeepBA), a graph neural network with only 2.04M parameters that recovers camera poses, intrinsics, and scene structure from 2D matches across an arbitrary number of views.
- Approach: amortizes the per-scene non-linear least-squares solve of classical SfM into a learned graph network that iteratively refines estimates from geometric residuals — converging from a cold start where classical methods need careful initialization.
- Training: scale-invariant supervision (ground-truth rotation, translation direction, intrinsics, plus reprojection self-consistency) lets datasets with arbitrary metric units be mixed without per-scene normalization.
- Results: outperforms prior learned SfM-from-matches on scene reconstruction and most pose metrics, is competitive with classical SfM at moderate thresholds and overtakes it in wall time as view count grows, and approaches monolithic models like VGGT (1.2B params) with greater modularity and matcher decoupling.
Related event: TUM's DeepBA: 2M-Parameter GNN Achieves Cold-Start Bundle Adjustment(2 posts)→
More from Research
- Research agenda: 'self-inoculation' to keep models' minds intact through RL training — repligate · 2026-11-18
- Embed team shows 3 years of fMRI-to-image BCI work, betting brain interfaces help align superintelligence — jonxuxu · 2026-11-18
- Robo-ValueRL rates actions by estimated task progress to learn from messy robot data — burkov · 2026-11-18
- RFD4-Proteina protein design GUI now runs in your browser, even on phones — AllThingsApx · 2026-11-18
- Redditor Asks: Can AI Now Formalize All of Math, CS, and Theoretical Physics in Lean? — SignificanceBulky162 · 2026-11-18
- Argon open-sources Alloy, a multi-material physics engine for robot learning on one GPU — ycombinator · 2026-11-18