Research map / 3D vision / Human pose estimation
iG-6DoF: Model-free 6DoF Pose Estimation for Unseen Object via Iterative 3D Gaussian Splatting
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 iG-6DoF, a model-free 6DoF pose estimator for unseen objects that combines multi-scale rotation-equivariant (icosahedral group) feature matching for initial pose with an iterative 3D Gaussian Splatting render-and-compare refiner.
Problem
Estimating 6DoF poses of unseen objects from reference images without CAD models or depth, robust to large translations/rotations and weak textures.
Approach
Augments reference images at multiple scales and rotations, matches query features in an icosahedral-group rotation-equivariant feature space to get an initial pose, then refines the pose by iteratively rendering a 3DGS object model and minimizing SSIM-based losses against the input image.
Key results
- On LINEMOD subset, refined iG-6DoF reaches 83.22% ADD(S)-0.1d vs 70.72% for Cas6D; without refinement it gets 45.99%.
- On OnePose-LowTexture (GT mask), iG-6DoF achieves 17.2 / 55.6 / 75.1 at 1cm-1deg / 3cm-3deg / 5cm-5deg, above OnePose++ (16.8 / 57.7 / 72.1) at the two strictest thresholds.
- On GenMOP, refined iG-6DoF gets average ADD-0.1d 66.83% vs 61.50% for Gen6D with reference refinement.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the 3DGS pose refiner (ablation) | iG-6DoF · LINEMOD subset | ADD(S)-0.1d: 45.99 -> 83.22 | Helped |
| replace the 3DGS refiner with the Gen6D refiner (ablation) | iG-6DoF · LM | ARV: 0.549 -> 0.538 | Hurt |
| replace the 3DGS refiner with the DeepIM refiner (ablation) | iG-6DoF · LM | ARV: 0.549 -> 0.512 | Hurt |
| replace the pose estimator with the ObjDesc selector (ablation) | iG-6DoF · LM | ARV: 0.549 -> 0.424 | Hurt |
| replace the pose estimator with the Gen6D selector (ablation) | iG-6DoF · LM | ARV: 0.549 -> 0.432 | Hurt |
| remove multi-scale data augmentation (ablation) | iG-6DoF · LM | ARV: 0.549 -> 0.521 | Hurt |
| use 16 reference images instead of the default count (ablation) | iG-6DoF · LM | ARV: 0.549 -> 0.432 | Hurt |
| use 32 reference images instead of the default count (ablation) | iG-6DoF · LM | ARV: 0.549 -> 0.446 | Hurt |
| use 64 reference images instead of the default count (ablation) | iG-6DoF · LM | ARV: 0.549 -> 0.533 | Hurt |
| use 128 reference images instead of the default count (ablation) | iG-6DoF · LM | ARV: 0.549 -> 0.587 | Helped |
| use iG-6DoF instead of MFOS | LINEMOD · LINEMOD | ADD(S)-0.1d: 78.4 -> 85.1 | Helped |
| use iG-6DoF instead of OnePose++ | OnePose-LowTexture · OnePose-LowTexture, 1cm-1deg | accuracy: 16.8 -> 17.2 | Helped |
Surprising findings
- On OnePose-LowTexture, the method without GT mask (Ours 16.6/53.2/73.5) is competitive with methods that use GT masks.
Limitations
- Refinement takes about 0.4 s per image, which limits speed relative to pure feed-forward methods.
- Paper notes that prior render-and-compare methods struggle to converge when rendered and query images differ substantially; the refiner's convergence under large disparities is not fully resolved.
- Some reported comparisons use different metrics or setups (e.g., BOP ablation vs. ADD tables), making direct comparison harder.
Experimental setup
Datasets: MegaPose synthetic dataset LINEMOD OnePose-LowTexture GenMOP
Benchmarks: LINEMOD OnePose-LowTexture GenMOP
Metrics: ADD(S)-0.1d ADD-AUC Proj-5 / Proj2D cm-degree accuracy BOP metric ARV/ARMSSD/ARMSPD
Baselines: Gen6D OnePose OnePose++ MFOS GS-Pose Cas6D OSOP LocPoseNet ObjDesc DeepIM PVNet RLLG
Builds on: 3D Gaussian Splatting RoReg Gen6D DeepIM TGID
Closest papers by meaning
- No Pose, No Problem: Surprisingly Simple 3D Gaussian Splats from Sparse Unposed Images (ICLR 2025 · similarity 0.89)
- Any6D: Model-free 6D Pose Estimation of Novel Objects (CVPR 2025 · similarity 0.89)
- Pos3R: 6D Pose Estimation for Unseen Objects Made Easy (CVPR 2025 · similarity 0.89)
- SelfSplat: Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting (CVPR 2025 · similarity 0.88)
- One2Any: One-Reference 6D Pose Estimation for Any Object (CVPR 2025 · similarity 0.88)
- KeyGS: A Keyframe-Centric Gaussian Splatting Method for Monocular Image Sequences (AAAI 2025 · 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.