TUM's DeepBA: 2M-Parameter GNN Achieves Cold-Start Bundle Adjustment

TUM's Daniel Cremers team released DeepBA, a 2.04M-parameter GNN that recovers camera poses from 2D matches across arbitrary views. About 590x smaller than VGGT, it surpasses existing learned SfM methods in cold-start bundle adjustment without requiring initialization.

2026-11-18 ~ 2026-11-18 · 2 related posts