Research map / 3D vision / Human pose estimation
One2Any: One-Reference 6D Pose Estimation for Any Object
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 One2Any, a single reference RGB-D view method that encodes the reference into a Reference Object Pose Embedding and decodes a Reference Object Coordinate map for query views via a latent-diffusion-based U-Net, with Umeyama alignment for 6D pose.
Problem
6D object pose estimation for novel objects when only a single reference RGB-D view is available, without CAD models, multi-view images, or category constraints.
Approach
Relaxes NOCS canonical-frame requirement by defining Reference Object Coordinates in the reference camera frame, and conditions a feedforward U-Net (initialized from Stable Diffusion) on a learned reference embedding (ROPE) via cross-attention to predict ROC maps for query images, then solves pose with Umeyama.
Key results
- On Real275 with a single reference, One2Any reaches 54.9 AR and 41.0 ADD-0.1d, above Oryon (46.5 AR, 34.9 ADD-0.1d).
- On occluded YCB-Video with one reference, One2Any gets 93.7 ADD-S AUC and 84.4 ADD AUC, beating Oryon and NOPE and outperforming several multi-view methods.
- Inference takes 0.09 s on an RTX 4090, more than 10x faster than Oryon (0.90 s), FoundationPose (2.70 s), and MegaPose (2.53 s).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use One2Any instead of Oryon | One2Any · Real275 | ADD-0.1d: 34.9 -> 41.0 | Helped |
| use One2Any instead of Oryon | One2Any · Toyota-Light | ADD-0.1d: 22.9 -> 34.6 | Helped |
| use One2Any instead of Oryon | One2Any · YCB-Video occluded (mean) | ADD AUC: 7.4 -> 84.4 | Helped |
| use One2Any instead of FoundationPose with 16 views | One2Any · YCB-Video occluded (mean) | ADD AUC: 91.5 -> 84.4 | Hurt |
| use One2Any instead of Oryon | One2Any · Inference | time (s): 0.90 -> 0.09 | Helped |
| replace ROC prediction with rotation and translation heads (ablation) | One2Any · YCB-Video ablation | ADD: 78.6 -> 59.4 | Hurt |
| add ROC to reference RGB and mask (ablation) | One2Any · YCB-Video ablation | ADD-S: 88.1 -> 91.2 | Helped |
| add reference RGB to ROC and mask (ablation) | One2Any · YCB-Video ablation | ADD-S: 90.5 -> 91.2 | Helped |
| replace reference ROC with depth (ablation) | One2Any · YCB-Video ablation | ADD-S: 91.2 -> 88.7 | Hurt |
| replace VQVAE query features with DINOv2 (ablation) | One2Any · YCB-Video ablation | ADD-S: 91.2 -> 91.6 | Helped |
| add depth to query RGB features (ablation) | One2Any · YCB-Video ablation | ADD-S: 91.2 -> 90.0 | Hurt |
Surprising findings
- Single-view One2Any outperforms FS6D (16 views, fine-tuned on YCB-Video) on occluded YCB-Video, and only FoundationPose with 16 views beats it.
- DINOv2 features give slightly better ablation numbers than the chosen frozen VQVAE features.
Limitations
- Trained with ground-truth depth and masks; relies on mask quality at test time (predicted masks only said to work comparably)
- Lags multi-view methods when those have enough views for good 3D reconstruction (e.g., FoundationPose-16view on YCB-Video and LINEMOD)
- Training required about 2 weeks on 8 RTX 4090 GPUs
Experimental setup
Datasets: FoundationPose synthetic data (Objaverse, GSO) OO3D-9D Real275 Toyota-Light LINEMOD YCB-Video T-LESS LINEMOD-Occlusion
Benchmarks: Real275 Toyota-Light YCB-Video LINEMOD
Metrics: AR ADD-0.1d ADD AUC ADD-S AUC 5°5cm 10°5cm
Baselines: Oryon NOPE ObjectMatch LatentFusion FoundationPose FS6D PoseDiffusion RelPose++ MegaPose GigaPose
Builds on: NOCS Stable Diffusion Oryon FoundationPose
Closest papers by meaning
- Any6D: Model-free 6D Pose Estimation of Novel Objects (CVPR 2025 · similarity 0.91)
- SingRef6D: Monocular Novel Object Pose Estimation with a Single RGB Reference (NeurIPS 2025 · similarity 0.89)
- UNOPose: Unseen Object Pose Estimation with an Unposed RGB-D Reference Image (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)
- Pos3R: 6D Pose Estimation for Unseen Objects Made Easy (CVPR 2025 · similarity 0.88)
- NOPE: Novel Object Pose Estimation from a Single Image (CVPR 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.