Research map / 3D vision / Human pose estimation
SingRef6D: Monocular Novel Object Pose Estimation with a Single RGB Reference
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 SingRef6D, a pose estimation pipeline that uses only a single RGB reference image, fine-tuning Depth-Anything v2 with a token scaler and depth-aware LoFTR matching.
Problem
Monocular 6D novel object pose estimation with only a single RGB reference image, where depth sensors fail on transparent/reflective surfaces and RGB matching fails in low-light or texture-less scenes.
Approach
A lightweight token scaler fine-tunes a frozen Depth-Anything v2 to predict metric depth on challenging surfaces (with scale, edge, and normal losses), and the predicted depth is fused into LoFTR's latent space for depth-aware matching, followed by PointDSC pose solving.
Key results
- On REAL275 depth prediction, the method improves delta_1.05 by 14.41% over Depth-Anything v2 (FT); on ClearPose transparent objects accuracy rises from 31.23% to 54.30%.
- Depth-aware LoFTR matching with predicted depth reaches 31.7 AR on Toyota-Light versus 24.1 for Oryon, and 28.7 AR on REAL275 versus 20.4 for Oryon with ground-truth depth settings in the table.
- Model has 11.6M parameters and 13.9 GFLOPs versus Oryon's 264.3M parameters and 120.1 GFLOPs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SingRef6D instead of Oryon with SingRef6D depth | 6D pose estimation · REAL275 | AR: 20.4 -> 28.7 | Helped |
| Use SingRef6D instead of RoMA with SingRef6D depth | 6D pose estimation · Tyo-L | AR: 30.9 -> 31.7 | Helped |
| Use SingRef6D instead of RoMA with SingRef6D depth | 6D pose estimation · ClearPose | AR: 17.7 -> 19.4 | Helped |
| Use token-scaler depth fine-tuning instead of Depth-Anything v2 (FT) | metric depth estimation · ClearPose | δ1.05: 31.23 -> 54.30 | Helped |
| Add scale, edge, and normal losses (ablation) | SingRef6D depth · REAL275 | δ1.05: 31.16 -> 44.28 | Helped |
| Add the scale loss to edge and normal losses (ablation) | SingRef6D depth · REAL275 | δ1.05: 40.23 -> 44.28 | Helped |
| Add the edge loss to scale and normal losses (ablation) | SingRef6D depth · REAL275 | δ1.05: 40.25 -> 44.28 | Helped |
| Add the normal loss to scale and edge losses (ablation) | SingRef6D depth · REAL275 | δ1.05: 40.41 -> 44.28 | Helped |
| Add the token scaler rather than fine-tuning only the depth head (ablation) | Depth-Anything v2 · REAL275 | δ1.05: 29.87 -> 44.28 | Helped |
| Add depth and latent fusion to the matcher (ablation) | LoFTR · Tyo-L | AR: 27.2 -> 31.7 | Helped |
| Add depth without latent fusion (ablation) | LoFTR · Tyo-L | AR: 27.2 -> 26.6 | Hurt |
| Use additive fusion instead of positional encoding for depth features (ablation) | LoFTR · Tyo-L | AR: 30.8 -> 31.7 | Helped |
Surprising findings
- On Toyota-Light, Oryon with DPAv2 depth outperforms the proposed method, which the authors attribute to depth prediction errors.
Limitations
- Relies on object masks to localize targets and constrain matching.
- Generalization bounded by Depth-Anything v2 and pretrained LoFTR; fails in extremely dark conditions.
- Toyota-Light performance trails Oryon with DPAv2 depth due to depth prediction errors.
Experimental setup
Datasets: REAL275 Toyota-Light ClearPose
Benchmarks: REAL275 Toyota-Light ClearPose
Metrics: delta 1.05 delta 1.10 delta 1.25 RMSE log10 Abs.Rel. Sq.Rel. MAE AR VSD
Baselines: Oryon SIFT RoMA UniDepth Depth-Anything v2 ScaleDepth Metric3D SPIDepth
Builds on: Depth-Anything v2 LoFTR PointDSC ControlNet
Closest papers by meaning
- RefPose: Leveraging Reference Geometric Correspondences for Accurate 6D Pose Estimation of Unseen Objects (CVPR 2025 · similarity 0.90)
- Pos3R: 6D Pose Estimation for Unseen Objects Made Easy (CVPR 2025 · similarity 0.89)
- Any6D: Model-free 6D Pose Estimation of Novel Objects (CVPR 2025 · similarity 0.89)
- MatchU: Matching Unseen Objects for 6D Pose Estimation from RGB-D Images (CVPR 2024 · similarity 0.89)
- One2Any: One-Reference 6D Pose Estimation for Any Object (CVPR 2025 · similarity 0.89)
- UNOPose: Unseen Object Pose Estimation with an Unposed RGB-D Reference Image (CVPR 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.