AI Paper Atlas

Research map / 3D vision / 3D generation

Curvature-Aware Captioning: Leveraging Geodesic Attention for 3D Scene Understanding

Ziyao He, Yingjie Liu, Zhang Yangrui, Mingsong Chen, Xuan Tang, Xian Wei

CVPR 2026 · 3D vision › 3D generation · approach: Semantic scene completion

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add Lorentz (hyperbolic) manifold alongside Oblique manifold (ablation)CAC · 3D dense captioning (IoU 0.50, MLE)CIDEr@0.5: 68.07 -> 69.92Improved
map decoder self-attention to Oblique manifold (ablation)CAC(O) · 3D dense captioning (IoU 0.50, MLE)CIDEr@0.5: 68.07 -> 62.96Hurt
map encoder and decoder both to Oblique manifold (ablation)CAC(O) · 3D dense captioning (IoU 0.50, SCST)CIDEr@0.5: 79.09 -> 70.40Hurt
map encoder to Hyperbolic instead of Oblique manifold (ablation)CAC · 3D dense captioning (IoU 0.50, SCST)CIDEr@0.5: 79.09 -> 77.46 CIDEr; BLEU-4 38.96 -> 39.47Mixed
use Version 1 bidirectional attention input configuration (ablation)BiCA · 3D dense captioning (IoU 0.50, MLE)CIDEr@0.5: 63.17 / 64.91 -> 65.22Improved
replace baseline attention with Version 1 input scheme (Version 4) (ablation)CAC(O) · 3D dense captioning (IoU 0.50, MLE)CIDEr@0.5: 68.07 -> 67.07Hurt
use Oblique plus Lorentz manifolds instead of Oblique only (ablation)CAC · Convergence speed (ScanRefer MLE)peak epoch: 337 -> 225Improved
use CAC(O&H) instead of Vote2Cap-DETR++3D dense captioning · Convergence speed (ScanRefer MLE)total time (s): 19,872.9 -> 19,667.7Improved
set Lorentz temperature to 0.1 and curvature c to 2 (ablation)CAC · Hyperparameter sensitivity (MLE)CIDEr@0.5: 50.03 -> 50.87Improved
lower curvature to c=0.5 (ablation)CAC · Hyperparameter sensitivity (MLE)CIDEr@0.5: 50.61 -> 49.90Hurt
set Oblique temperature to 0.5 instead of 1 in pretraining (ablation)CAC pretraining · Pretraining detection (epoch 119)mAP@0.50: 38.05 -> 36.43Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy