Research map / 3D vision / Human pose estimation
Co-op: Correspondence-based Novel Object 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.
Co-op estimates semi-dense patch-level correspondences between a query image and a few CAD-rendered templates for coarse pose, then refines with a probabilistic-flow render-and-compare model and differentiable PnP, achieving state-of-the-art unseen-object pose on BOP's seven core datasets without per-object training.
Problem
Estimating 6DoF poses of objects unseen during training from a single RGB image using only CAD models, while remaining fast and accurate without the large template sets used by prior model-based methods.
Approach
Coarse estimation uses a hybrid of patch-level classification and offset regression to produce semi-dense correspondences to only out-of-plane-rotation templates (in-plane invariant), and refinement learns a Laplace-distributed probabilistic flow whose confidence weights a differentiable PnP/LM pose update.
Key results
- Coarse-only Co-op reaches mean AR 58.4 vs 37.3 for FoundPose (second best coarse method) on BOP's seven core datasets.
- With refinement and 5-hypothesis pose selection, Co-op reaches mean AR 65.7 vs 59.6 for FoundPose+MegaPose.
- Ablation: removing patch-level classification drops coarse mean AR from 58.4 to 50.2; removing probabilistic flow drops refined mean AR from 64.0 to 63.3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Co-op coarse estimation instead of FoundPose coarse estimation | Co-op · BOP core datasets (mean) | AR: 37.3 -> 58.4 | Helped |
| use the Co-op refiner with Co-op coarse estimation instead of GigaPose with the Co-op refiner | Co-op · BOP core datasets (mean) | AR: 58.3 -> 64.0 | Helped |
| use Co-op with five hypotheses instead of FoundPose with five hypotheses | Co-op · BOP core datasets (mean) | AR: 59.6 -> 65.7 | Helped |
| add patch-level classification to offset regression (ablation) | Co-op coarse estimation · BOP core datasets | AR: 50.2 -> 58.4 | Helped |
| add CroCo pretraining (ablation) | Co-op coarse estimation · BOP core datasets | AR: 46.4 -> 58.4 | Helped |
| use probabilistic flow regression (ablation) | Co-op pose refinement · BOP core datasets | AR: 63.3 -> 64.0 | Helped |
| add CroCo pretraining (ablation) | Co-op pose refinement · BOP core datasets | AR: 61.2 -> 64.0 | Helped |
| add CroCo pretraining (ablation) | Co-op pose selection · BOP core datasets | AR: 65.0 -> 65.7 | Helped |
Limitations
- Multiple-hypothesis pose selection increases inference time (about 4.2 s per image with 5 hypotheses).
- Relies on CNOS segmentation masks for object detection in BOP evaluation.
Experimental setup
Datasets: MegaPose synthetic dataset BOP LM-O T-LESS TUD-L IC-BIN ITODD HB YCB-V
Benchmarks: BOP LM-O T-LESS TUD-L IC-BIN ITODD HB YCB-V
Metrics: Average Recall (AR) VSD recall MSSD recall MSPD recall inference time
Baselines: ZS6D MegaPose GenFlow GigaPose FoundPose
Builds on: GigaPose GenFlow CroCo PDC-Net+ CNOS
Closest papers by meaning
- UNOPose: Unseen Object Pose Estimation with an Unposed RGB-D Reference Image (CVPR 2025 · similarity 0.89)
- GigaPose: Fast and Robust Novel Object Pose Estimation via One Correspondence (CVPR 2024 · similarity 0.89)
- RefPose: Leveraging Reference Geometric Correspondences for Accurate 6D Pose Estimation of Unseen Objects (CVPR 2025 · similarity 0.89)
- Pos3R: 6D Pose Estimation for Unseen Objects Made Easy (CVPR 2025 · similarity 0.88)
- GIVEPose: Gradual Intra-class Variation Elimination for RGB-based Category-Level Object Pose Estimation (CVPR 2025 · similarity 0.88)
- MFOS: Model-Free & One-Shot Object Pose Estimation (AAAI 2024 · similarity 0.88)
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.