MVLoMa extends LoMa to multi-view feature matching with geometry foundation models, leading 25-view benchmarks
zhenjun_zhao · x · 2026-10-24
MVLoMa extends the two-view matcher LoMa to multi-view: a frozen multi-view geometry foundation model (DA3-G etc.) supplies geometry-aware patch features, fused with trainable per-view VGG local features via a convolutional refiner, then a multi-view star-graph matcher jointly processes the centre image and all neighbours.
- Averages 58.65/85.41/90.99 AUC@1°/5°/10° across HyperSim, ScanNet++, Mip-NeRF 360 and Tanks and Temples at 25 views, beating MV-RoMa (OOM beyond 13 views) and two-view methods
- Ablations show DA3-G is the best frozen encoder, ahead of π³ and VGGT variants
More from Research
- Rice & NVIDIA's Sensorimotor Alignment predicts robot policy success without training (r=0.717) — CSProfKGD · 2026-10-24
- RLVR Trains VLMs on Instruction Hierarchy with Just 7K Verifiable Examples — TuhinChakr · 2026-10-24
- NVIDIA ships BioNeMo Inference Runtime v0.2 with RFD4-Proteina support — AllThingsApx · 2026-10-24
- Trillion-observation finance benchmark: linear ridge probe beats 11 of 14 self-supervised encoders — burny_tech · 2026-10-24
- Nvidia and UW Institute for Protein Design open-source RFD4-Proteina protein design model — AllThingsApx · 2026-10-24
- Ben Goertzel argues the selfish-supermind attractor has a wider basin than BGI — bengoertzel · 2026-10-24