Research map / 3D vision / Human pose estimation
GIVEPose: Gradual Intra-class Variation Elimination for RGB-based Category-Level 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.
Proposes GIVEPose, an RGB-only category-level pose estimator that predicts a NOCS map and then uses a deformable convolutional autoencoder to convert it into an intra-class-variation-free consensus (IVFC) map for geometry-guided pose regression.
Problem
RGB-only category-level object pose estimation via geometry-guided regression is hampered because the NOCS map carries instance-specific intra-class shape variation, creating a many-to-one mapping to category-level pose.
Approach
Introduces the IVFC map, rendered from a fixed category-consensus model so it has a one-to-one relation with category pose, and gradually eliminates intra-class variation by mapping the NOCS map to the IVFC map with a deformable convolutional autoencoder fused with backbone features.
Key results
- On REAL275, GIVEPose reaches NIOU75 20.8 and 10°0.2d 44.6, surpassing LaPose by 5.0 and 7.2 points.
- On CAMERA25, NIOU75 improves to 19.4 from LaPose's 14.1 and 10°0.2d to 47.5.
- On Wild6D, NIOU75 is 28.1 versus LaPose's 24.4 when trained on CAMERA25 and REAL275.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add NOCS map supervision (ablation) | GIVEPose · REAL275 | NIoU75: 18.2 -> 20.8 | Helped |
| add IVFC map supervision (ablation) | GIVEPose · REAL275 | NIoU75: 16.4 -> 20.8 | Helped |
| use gradual variation elimination instead of parallel map prediction and concatenation (ablation) | GIVEPose · REAL275 | NIoU75: 18.0 -> 20.8 | Helped |
| use DCAE instead of standard CAE (ablation) | GIVEPose · REAL275 | NIoU75: 17.5 -> 20.8 | Helped |
| use DCAE instead of ViT-AE (ablation) | GIVEPose · REAL275 | NIoU75: 18.3 -> 20.8 | Helped |
| use IVFC map instead of Ball map (ablation) | GIVEPose · REAL275 | NIoU75: 17.8 -> 20.8 | Helped |
Surprising findings
- Using NOCS map or IVFC map alone as intermediate representation yields similar pose accuracy, despite IVFC map being easier to regress from 2D-3D correspondences with ground truth.
Limitations
- Training relies on expensive annotated real-world data.
- Overall accuracy is constrained by the independent detector's performance.
- Operates on cropped images from an off-the-shelf detector rather than end-to-end.
Experimental setup
Datasets: CAMERA25 REAL275 Wild6D
Benchmarks: REAL275 CAMERA25 Wild6D
Metrics: NIOU25 NIOU50 NIOU75 10°0.2d 10°0.5d IoU25 IoU50 10°10cm mAP
Baselines: MSOS OLD-Net DMSR LaPose
Builds on: LaPose NOCS Deformable convolution Denoising autoencoder MaskRCNN
Closest papers by meaning
- SchemaPose: RGB-Based Category-Level Pose Estimation with Parametric Category Schema (NeurIPS 2026 · similarity 0.91)
- Co-op: Correspondence-based Novel Object Pose Estimation (CVPR 2025 · similarity 0.88)
- CAP-Net: A Unified Network for 6D Pose and Size Estimation of Categorical Articulated Parts from a Single RGB-D Image (CVPR 2025 · similarity 0.88)
- Universal Features Guided Zero-Shot Category-Level Object Pose Estimation (AAAI 2025 · similarity 0.87)
- Rethinking Correspondence-based Category-Level Object Pose Estimation (CVPR 2025 · similarity 0.87)
- GeoReF: Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement (CVPR 2024 · similarity 0.87)
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.