Research map / 3D vision / Human pose estimation
UNOPose: Unseen Object Pose Estimation with an Unposed RGB-D Reference Image
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.
Estimates the 6DoF relative pose of an unseen object from a single unposed RGB-D reference image using an SE(3)-invariant global/local reference frame coarse-to-fine network with an overlap predictor, evaluated on a new BOP-derived one-reference benchmark.
Problem
Pose estimation of unseen objects when only one unposed RGB-D reference image is available, where the relative pose spans all of SE(3) and partial viewpoint overlap is low.
Approach
Builds an SE(3)-invariant global reference frame (from PCA normals and centroid) to canonicalize pose and size, adds local reference frame encodings for fine matching, and reweights correspondences using a predicted overlap-region confidence within a coarse-to-fine geometric transformer registration.
Key results
- UNOPose reaches 70.9% average ARBOP across LM-O, TUD-L and YCB-V, outperforming image-reference baselines such as FreeZe (59.8%) and SAM-6D with posed reference (50.8%).
- UNOPose is competitive with CAD-based ZTE-PPF (69.0%) but below Koenig-PPF (75.1%) on the same benchmark.
- UNOSeg segmentation reaches 54.2% mAP average, close to SAM-6D's 54.5% while running faster (2.68s vs 4.53s).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UNOPose with one unposed RGB-D reference | Unseen object pose estimation · LM-O, TUD-L, and YCB-V | ARBOP: 58.7; 71.0; 83.1 | Helped |
| Add the global reference frame (GRF) (ablation) | UNOPose coarse pose estimation · YCB-V | ARBOP: 63.9 -> 67.5 | Helped |
| Add the local reference frame (LRF) (ablation) | UNOPose coarse pose estimation · YCB-V | ARBOP: 63.9 -> 65.4 | Helped |
| Add fine pose estimation to the model with GRF and LRF (ablation) | UNOPose · YCB-V | ARBOP: 67.9 -> 81.5 | Helped |
| Add GRF to the fine-pose model without LRF (ablation) | UNOPose · YCB-V | ARBOP: 80.1 -> 80.4 | Helped |
| Add LRF to the fine-pose model without GRF (ablation) | UNOPose · YCB-V | ARBOP: 80.1 -> 80.9 | Helped |
| Add the overlap predictor instead of soft correspondence (ablation) | UNOPose · YCB-V | ARBOP: 77.2 -> 83.1 | Helped |
| Add the overlap predictor to coarse pose estimation (ablation) | UNOPose coarse pose estimation · YCB-V | ARBOP: 67.9 -> 69.2 | Helped |
| Replace UNOSeg with CNOS (ablation) | UNOPose · YCB-V | ARBOP: 83.1 -> 77.3 | Hurt |
| Replace UNOSeg with SAM-6D segmentation (ablation) | UNOPose · YCB-V | ARBOP: 83.1 -> 78.2 | Hurt |
| Replace UNOSeg with MRCNN (ablation) | UNOPose · YCB-V | ARBOP: 83.1 -> 81.6 | Hurt |
| Use ground-truth segmentation instead of UNOSeg (ablation) | UNOPose · YCB-V | ARBOP: 83.1 -> 86.0 | Helped |
Surprising findings
- Using a rendered (pose-identical) reference instead of the real reference improves results (85.5% vs 77.2% overlap-predictor baseline... specifically 85.5 vs 83.1 ARBOP on YCB-V), suggesting depth noise and domain gap hurt performance.
- Soft correspondence loss from prior work underperforms the proposed overlap predictor (77.2 vs 83.1 ARBOP on YCB-V).
Limitations
- Correspondence-based approach depends on robust dense correspondences, limiting extreme two-view geometry cases.
- Performance drops sharply when initial rotation distance between query and reference exceeds 50 degrees due to low overlap.
- Requires an RGB-D reference with reasonable visibility (occlusion under 30%).
Experimental setup
Datasets: MegaPose synthetic dataset GSO ShapeNet
Benchmarks: LM-O TUD-L YCB-V
Metrics: mAP ARBOP (average of VSD, MSSD, MSPD recall) run time
Baselines: ZTE-PPF Koenig-PPF SAM-6D FreeZe GeDi UTOPIC FCGF+RANSAC FPFH+RANSAC PPF_3D_ICP ZeroPose CNOS
Builds on: GeoTransformer SAM-6D DINOv2 InfoNCE MegaPose
Closest papers by meaning
- RefPose: Leveraging Reference Geometric Correspondences for Accurate 6D Pose Estimation of Unseen Objects (CVPR 2025 · similarity 0.91)
- 3D-Aware Hypothesis & Verification for Generalizable Relative Object Pose Estimation (ICLR 2024 · similarity 0.90)
- Co-op: Correspondence-based Novel Object Pose Estimation (CVPR 2025 · similarity 0.89)
- UA-Pose: Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References (CVPR 2025 · similarity 0.89)
- NOPE: Novel Object Pose Estimation from a Single Image (CVPR 2024 · similarity 0.89)
- Universal Features Guided Zero-Shot Category-Level Object Pose Estimation (AAAI 2025 · similarity 0.89)
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.