ALoMa: Frobenius norm is a bad loss for affine correspondences, Charbonnier cuts error 3x
ducha_aiki · x · 2026-11-17
The ALoMa paper (Affine Correspondences Beyond Patches) predicts affine shapes from LoMa (LightGlue) descriptors, improving on LoMa. Author duchaaiki shares a practical lesson: he had to switch to a descriptor loss because the Frobenius norm is a bad loss.
Per the paper's tables, a robust Charbonnier loss (δ=10⁻³) instead of plain squared Frobenius cuts median affine error on MegaDepth-1500 from 11.9 to 4.3 (4.0 with better parametrization) — a 3x improvement.
Related event: ALoMa Paper Extends Affine Correspondence Estimation Beyond Patches(2 posts)→
More from Research
- RL agent navigates SMEFT operator space for new physics fits — Universite-de-Montreal · 2026-11-17
- EACL 2027 Paper Finds up to 74% Logical Error Rate in LLM Code Across 10 Models — ponguru · 2026-11-17
- AI agents rewrote 30,000 articles of French law; blind audits found 54% fewer legal inconsistencies — IgorCarron · 2026-11-17
- How rollout staleness hurts on-policy distillation, and how to build a faster pipeline — burkov · 2026-11-17
- Apple robotics paper: 91% success in unfamiliar homes using synthetic sim data — FuSheng_0306 · 2026-11-17
- Hierarchos-Native v0.2.0: Rust+Vulkan LLM training adds RDNA3 WMMA and native multimodal training — PhysicsDisastrous462 · 2026-11-17