Research map / 3D vision / Human pose estimation
Towards Co-Evaluation of Cameras HDR and Algorithms for Industrial-Grade 6DoF Pose Estimation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces the Industrial Plenoptic Dataset (IPD) of 20 industrial parts captured by 13 cameras under varied lighting and exposures, plus a Robot Consistency evaluation method for sub-3mm 6DoF pose accuracy.
Problem
6DoF pose estimation lacks benchmarks and evaluation methods that measure industrial-grade (sub-millimeter) accuracy and jointly account for cameras, HDR, lighting, and polarization; annotated ground truth is too inaccurate at 150-200cm working distances.
Approach
Evaluate pose estimators by robot consistency: a calibrated robot moves a rigidly mounted object through many configurations, and predicted poses are compared against the robot-derived consensus pose (Fréchet mean) instead of noisy human annotations. Paired with a multi-camera, multi-modal, multi-exposure, multi-lighting dataset.
Key results
- Basler-HR camera gives the most robust performance across lighting conditions, e.g. MVD 2.640 mm and recall 0.83 under roomlight versus 5.676 mm for Photoneo.
- Mertens HDR fusion gives the best median MVD under spotlight (3.061 mm, recall 0.77) among tested HDR methods, versus 3.271 mm with no HDR.
- Polarization (AOLP/DOLP) edge refinement reduces median MVD for FLIR-MonoP on pipe fitting from 9.92 mm to 8.43 mm and on oblong float from 6.17 mm to 5.41 mm.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use a high-resolution multi-view camera setup (Basler HR) instead of the other evaluated cameras | pose estimation pipeline · IPD, daylight, roomlight, and spotlight | MVD: 3.078; 2.640; 3.061 | Helped |
| use Mertens HDR instead of no HDR (ablation) | Basler-HR pose estimation pipeline · IPD, spotlight | Median MVD: 3.271 -> 3.061 | Helped |
| use Mertens HDR instead of no HDR (ablation) | Basler-HR pose estimation pipeline · IPD, spotlight | Recall: 0.72 -> 0.77 | Helped |
| use polarization (AOLP/DOLP) for edge refinement instead of no refinement modality (ablation) | FLIR-MonoP pose estimation pipeline · Pipe Fitting | Median MVD: 9.92 -> 8.43 | Helped |
| use polarization (AOLP/DOLP) for edge refinement instead of no refinement modality (ablation) | FLIR-MonoP pose estimation pipeline · Oblong Float | Median MVD: 6.17 -> 5.41 | Helped |
| evaluate poses with Robot Consistency instead of 1 mm-jitter annotations | pose-estimator evaluation · synthetic scenes | mean absolute difference (mm): 0.13 for our method vs. 0.47 mm for 1 mm jitter | Helped |
Surprising findings
- Photoneo structured light performs very poorly on dark, reflective objects (recall 0.06 on oblong float) and under spotlight ambient interference.
- Annotation-based evaluation systematically overestimates model error, with the bias worsening as true performance improves.
Limitations
- Robot Consistency theory does not account for robot or calibration errors and many forms of bias.
- Reference annotations require semi-automated clustering plus manual inspection and filtering.
- Dataset covers only 20 mostly mechanical parts with moderate clutter; all current methods perform poorly on thin, small, and dark reflective parts.
Experimental setup
Datasets: Industrial Plenoptic Dataset (IPD)
Benchmarks: Industrial Plenoptic Dataset (IPD)
Metrics: MVD recall precision ADD correlation with true performance
Baselines: Debevec HDR Robertson HDR Mertens HDR Basler-LR Basler-HR FLIR-MonoP Photoneo
Builds on: ADD metric Mask-RCNN Mertens exposure fusion edge refinement (Choi et al.) robot consistency (Kalra et al. prior work)
Closest papers by meaning
- SingRef6D: Monocular Novel Object Pose Estimation with a Single RGB Reference (NeurIPS 2025 · similarity 0.85)
- HouseCat6D - A Large-Scale Multi-Modal Category Level 6D Object Perception Dataset with Household Objects in Realistic Scenarios (CVPR 2024 · similarity 0.84)
- Real-IAD D3: A Real-World 2D/Pseudo-3D/3D Dataset for Industrial Anomaly Detection (CVPR 2025 · similarity 0.82)
- CAP-Net: A Unified Network for 6D Pose and Size Estimation of Categorical Articulated Parts from a Single RGB-D Image (CVPR 2025 · similarity 0.82)
- Any6D: Model-free 6D Pose Estimation of Novel Objects (CVPR 2025 · similarity 0.82)
- 6D Object Pose Tracking in Internet Videos for Robotic Manipulation (ICLR 2025 · similarity 0.82)
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.