Research map / 3D vision / Human pose estimation
Deep Bingham Networks: Dealing with Uncertainty and Ambiguity in Pose Estimation
Haowen Deng, Mai Bui, Nassir Navab, Leonidas J. Guibas, Slobodan Ilić, Tolga Birdal
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 Bingham-distribution loss (UBN) instead of L1 loss | quaternion regression with PointNet · ModelNet10 | Chamfer distance (x10^2): 3.605 -> 1.477 | Improved |
| replace unimodal UBN by mixture of 10 Bingham components (MBN-10) (ablation) | UBN · ModelNet10 | Chamfer distance (x10^2): 1.477 -> 0.838 | Improved |
| increase components from 10 to 25 and 50 (ablation) | MBN · ModelNet10 | Chamfer distance (x10^2): 0.838 -> 0.926 -> 0.973 | Hurt |
| add explicit mixture coefficient learning and multi-hypothesis training (MBN-CE) instead of finite mixture MBN-MB (ablation) | MBN-MB · Ambiguous relocalization | ratio of correct poses (15deg/0.2m): 0.28 -> 0.56 | Improved |
| implicit mixture coefficient learning (MBN) versus MBN-CE with 10 hypotheses (ablation) | MBN-CE · Ambiguous relocalization | ratio of correct poses (15deg/0.2m): 0.56 -> 0.58 | Improved |
| use multiple-hypothesis Bingham mixture instead of MC-Dropout | MC-Dropout · Synthetic round table | correctly detected modes: 24.8% -> 99.1% | Improved |
| use multiple-hypothesis Bingham mixture instead of MC-Dropout | MC-Dropout · Synthetic dining table | correctly detected modes: 50% -> 96% | Improved |
| choose best branch by l1 distance rather than probability (ablation) | MBN training · Ambiguous relocalization | ratio of correct poses | Improved |
| vary backbone (ResNet-18, 34, 50, Inception-v3) (ablation) | MBN · Ambiguous relocalization | ratio of correct poses (15deg/0.2m): 0.53 -> 0.57 -> 0.52 -> 0.50 | No clear effect |
| filter hypotheses by predicted uncertainty before registration (ablation) | registration · point cloud registration | average recall: 0.777 -> 0.730 | No clear effect |
| unimodal UBN versus single-pose regressors on non-ambiguous scenes | UBN · 7-Scenes Chess | median rotation error: 4.48 -> 4.97 | No clear effect |
Closest papers by meaning
- A Simple yet Effective Baseline for 3D Human Pose Estimation (ICCV 2017 · similarity 0.88)
- 6D Camera Relocalization in Ambiguous Scenes via Continuous Multimodal Inference (ECCV 2020 · similarity 0.88)
- Wide-Baseline Relative Camera Pose Estimation With Directional Learning (CVPR 2021 · similarity 0.87)
- Generating Multiple Hypotheses for 3D Human Pose Estimation With Mixture Density Network (CVPR 2019 · similarity 0.87)
- Exploiting temporal information for 3D human pose estimation (ECCV 2018 · similarity 0.87)
- LEAP: Liberate Sparse-View 3D Modeling from Camera Poses (ICLR 2024 · similarity 0.86)
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.
