SUFLECA shows NOC-based correspondence can improve CAD-to-image alignment
ducha_aiki · x · 2026-07-21
SUFLECA (“Scaling Up Feature Learning for CAD-to-image Alignment”) proposes a new way to train CAD-to-image alignment models.
The post’s TL;DR is that normalized object surface coordinates (NOCs) are a good proxy target for correspondence learning. The figures show the architecture: multi-scale features from a frozen perception encoder are fused with a Dense Prediction Transformer, then decoded by a lightweight binned NOC head to predict correspondences.
The paper also lists dataset composition and ablations:
- Training data spans several real and synthetic sources, with an overall total of 674,851 frames and 3.38 objects/frame.
- On ScanNet25k, the authors report 33.4 / 42.3 for category / instance alignment accuracy for SUFLECA.
- Ablations show that including ScanNet during training improves results to 34.4 / 42.8, while alternative correspondence or NMS scoring methods perform worse.
- A retrieval-accuracy table shows ground-truth CAD annotations at 100.0 retrieval accuracy, compared with 34.3 for ROCA and 3.8 for GroundedSAM + OSCAR, with corresponding drops in alignment accuracy.
Overall, the post argues that better correspondence modeling and CAD retrieval accuracy are central to alignment quality.
More from Research
- A new RLHF book is finished after nights and weekends since 2024 — HamelHusain · 2026-07-21
- Google DeepMind’s Project Genie talk shows how creatives feed into model research — alexanderchen · 2026-07-21
- Nat Lambert says RL distillation does not use the strongest models as teachers — natolambert · 2026-07-21
- Thread claims GPT-5.6 Sol helped build a new counterexample factory for the Jacobian conjecture — LucaAmb · 2026-07-21
- TimeLens2 claims SOTA on 7 video grounding benchmarks with 4B and 8B models — _akhaliq · 2026-07-21
- Real AI progress in math often comes from counterexamples to old beliefs — LucaAmb · 2026-07-21