Research map / 3D vision / Human pose estimation
GenFlow: Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects
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.
Proposes GenFlow, an optical-flow-based recurrent refiner with a differentiable PnP layer and shape-constrained correlation lookup that refines 6D poses of novel objects from RGB or RGB-D input.
Problem
Generalizable 6D object pose refinement for novel (unseen) objects, where prior refiners use the 3D shape only indirectly and trade accuracy for scalability.
Approach
Iteratively estimate 2D optical flow between the rendered and observed image, lift flow to 2D-3D correspondences via a differentiable PnP layer, and look up correlations using pose-induced flow so the network is constrained by the target's 3D shape; add a cascade of GenFlow modules over multi-scale correlation volumes and factorize confidence into certainty and pose sensitivity.
Key results
- On BOP unseen-object RGB-D setting with CNOS detection, Ours+Kabsch+MH reaches mean AR 67.2 versus 62.8 for MegaPose+MH+Teaserpp.
- On unseen objects with RGB input, Ours+MH achieves mean AR 57.8 versus 54.9 for MegaPose+MH using CNOS detection.
- Ablation: shape-constraint plus confidence factorization plus cascade improves mean AR on 5 BOP datasets from 66.3 (no shape constraint) to 70.7.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace MegaPose+MH with GenFlow+MH | Unseen-object RGB pose estimation · BOP datasets, mean across 7 datasets | Average Recall: 54.9 -> 57.8 | Helped |
| Replace MegaPose+MH+Teaserpp with GenFlow+Kabsch+MH | Unseen-object RGB-D pose estimation · BOP datasets, mean across 7 datasets | Average Recall: 62.8 -> 67.2 | Helped |
| Use GMM-based pose hypothesis sampling instead of naïve sampling (ablation) | GenFlow coarse estimator · BOP datasets (LM-O, T-LESS, TUD-L, IC-BIN, YCB-V) | Average Recall: 22.5 -> 31.3; 22.5 -> 35.8 | Helped |
| Use GMM-based sampling with 208 rather than 144 renderings (ablation) | GenFlow coarse estimator · BOP datasets (LM-O, T-LESS, TUD-L, IC-BIN, YCB-V) | Average Recall: 31.3 -> 35.8 | Helped |
| Add shape constraint to the refiner (ablation) | GenFlow refiner · BOP datasets (LM-O, T-LESS, TUD-L, IC-BIN, YCB-V) | Average Recall: 66.3 -> 69.8 | Helped |
| Add cascade architecture to the shape-constrained refiner (ablation) | GenFlow refiner · BOP datasets (LM-O, T-LESS, TUD-L, IC-BIN, YCB-V) | Average Recall: 69.8 -> 70.4 | Helped |
| Add confidence factorization to the shape-constrained refiner (ablation) | GenFlow refiner · BOP datasets (LM-O, T-LESS, TUD-L, IC-BIN, YCB-V) | Average Recall: 69.8 -> 69.9 | Helped |
| Add cascade architecture to the shape-constrained refiner with confidence factorization (ablation) | GenFlow refiner · BOP datasets (LM-O, T-LESS, TUD-L, IC-BIN, YCB-V) | Average Recall: 69.9 -> 70.7 | Helped |
Surprising findings
- A single GenFlow outer update is slower than the MegaPose refiner (98.2 ms vs 66.5 ms), yet it reaches better accuracy with the same or fewer iterations.
Limitations
- Single outer update is slower than the MegaPose refiner in wall-clock time.
- Novel-object detection is out of scope; results depend on the CNOS detector, and the best seen-object results still come from other methods.
- RGB-D refinement requires an additional RANSAC-Kabsch step.
Experimental setup
Datasets: MegaPose synthetic dataset ShapeNet Google Scanned Objects LM-O T-LESS TUD-L IC-BIN ITODD HomebrewDB YCB-V
Benchmarks: BOP LM-O BOP T-LESS BOP TUD-L BOP IC-BIN BOP ITODD BOP HB BOP YCB-V
Metrics: average recall (AR) VSD MSSD MSPD
Baselines: CosyPose SurfEmb ZebraPose MegaPose WDR-Pose PFA OSOP ZeroPose CNOS
Builds on: RAFT SCFlow PFA MegaPose CIR
Closest papers by meaning
- SCFlow2: Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow (CVPR 2025 · similarity 0.90)
- RefPose: Leveraging Reference Geometric Correspondences for Accurate 6D Pose Estimation of Unseen Objects (CVPR 2025 · similarity 0.86)
- FoundationPose: Unified 6D Pose Estimation and Tracking of Novel Objects (CVPR 2024 · similarity 0.86)
- SingRef6D: Monocular Novel Object Pose Estimation with a Single RGB Reference (NeurIPS 2025 · similarity 0.86)
- Rectified Point Flow: Generic Point Cloud Pose Estimation (NeurIPS 2025 · similarity 0.86)
- Co-op: Correspondence-based Novel Object Pose Estimation (CVPR 2025 · similarity 0.85)
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.