[HUGGINGFACE]score: 0.42
BA-T Applies Bundle Adjustment Iteration Structure to Transformer 3D Reconstruction
June 1, 2026
BA-T replaces deep decoder stacks in feed-forward 3D reconstruction models with a single lightweight repeatable transformer layer that performs bundle-adjustment-style iterative refinement in latent token space. The approach improves multi-view consistency without the compute overhead of heavy cross-view attention decoders.
paper
HOW THIS AFFECTS YOU
●
builderLighter inference head with better multi-view consistency could be relevant for real-time 3D reconstruction pipelines if benchmark numbers hold up.
●
researcherThe structured iterative refinement via latent residuals is the key architectural contribution — directly comparable to DUSt3R/MASt3R-style feed-forward reconstruction baselines.