Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Estimating and Exploiting the Aleatoric Uncertainty in Surface Normal Estimation
Gwangbin Bae, Ignas Budvytis, Roberto Cipolla
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 changed | Where | Measured | Effect |
|---|---|---|---|
| use angular vonMF NLL instead of vonMF NLL (ablation) | baseline encoder-decoder · ScanNet (20% train subset, ablation) | mean angular error: 14.10 -> 13.82 | Improved |
| add learned attenuation to the angular loss (AL to NLL-AngMF) (ablation) | baseline encoder-decoder · ScanNet (20% train subset, ablation) | mean angular error: 13.45 -> 13.82 mean; 6.70 -> 6.60 median | Mixed |
| add pixel-wise MLP decoder modules (ablation) | NLL-AngMF · ScanNet (20% train subset, ablation) | mean angular error: 13.82 -> 13.59 | Improved |
| add uncertainty-guided sampling to MLP training (ablation) | NLL-AngMF + MLPs · ScanNet (20% train subset, ablation) | mean angular error: 13.59 -> 13.17 | Improved |
| importance-sample uncertain pixels (beta_UG 0.0 to 0.7) (ablation) | full method · ScanNet (20% train subset, ablation) | mean angular error: 13.58 -> 13.17 | Improved |
| sample only highest-uncertainty pixels (beta_UG 1.0 vs 0.7) (ablation) | full method · ScanNet (20% train subset, ablation) | mean angular error: 13.17 -> 13.26 | Hurt |
| use the proposed method versus GeoNet++ | Ours · NYUv2 (trained on NYUv2) | mean angular error: 18.5 -> 14.9 | Improved |
| use the proposed method versus TiltedSN | Ours · ScanNet (full train) | mean angular error: 12.6 -> 11.8 | Improved |
| cross-dataset transfer ScanNet to NYUv2 versus TiltedSN | Ours · NYUv2 (trained on ScanNet) | accuracy within 11.25 deg: 59.8 -> 59.0 | No clear effect |
| use predicted NLL uncertainty instead of dropout plus augmentation | Ours (NLL-AngMF) · NYUv2 uncertainty | AUSC mean: 8.16 -> 6.83 | Improved |
| AngMF instead of vonMF for uncertainty (ablation) | Ours · NYUv2 uncertainty | AUSC mean / AUSE mean: 7.03 -> 6.83 AUSC; 2.11 -> 2.13 AUSE | Mixed |
| combine test-time dropout and augmentation (ablation) | Drop · NYUv2 uncertainty | AUSC mean: 9.01 -> 8.16 | Improved |
Closest papers by meaning
- Rethinking Inductive Biases for Surface Normal Estimation (CVPR 2024 · similarity 0.87)
- Fine-Scale Surface Normal Estimation Using a Single NIR Image (ECCV 2016 · similarity 0.87)
- Deep Iterative Surface Normal Estimation (CVPR 2020 · similarity 0.87)
- Adaptive Surface Normal Constraint for Geometric Estimation From Monocular Images (TPAMI 2024 · similarity 0.87)
- Surface Normal Estimation of Tilted Images via Spatial Rectifier (ECCV 2020 · similarity 0.86)
- Variational Multi-scale Representation for Estimating Uncertainty in 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.84)
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.
