Research map / 3D vision / Human pose estimation
GCE-Pose: Global Context Enhancement for 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 GCE-Pose, which reconstructs a complete semantic shape from partial RGB-D via a deep linear shape model over category DINOv2 prototypes and fuses it with keypoint features to improve model-free category-level pose estimation.
Problem
Category-level object pose estimation from partial RGB-D observations, where partial visibility limits extraction of global context features that generalize across instances in a category.
Approach
A first-complete-then-aggregate strategy: a per-category deep linear semantic shape model deforms a mean shape with DINOv2 semantic prototypes to reconstruct the global geometry and semantics of a partial instance, and a cross-attention GCE module fuses this global context into keypoint features before NOCS and pose regression.
Key results
- On HouseCat6D, GCE-Pose reaches 24.8 on 5°2cm and 79.2 IoU50 vs 21.3 and 76.9 for AG-Pose (DINO).
- On NOCS-REAL275, GCE-Pose reaches 57.0 on 5°2cm and 79.8 IoU75, the best on most metrics versus AG-Pose (DINO) at 57.0 and 80.1 IoU75 (a tie on 5°2cm, slightly lower IoU75).
- Ablation on HouseCat6D: mean-shape geometry plus categorical semantics gives 23.4 on 5°2cm, and the full pipeline 24.8, versus 21.3 for the baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add GCE-Pose global context priors and feature fusion | AG-Pose (DINO) · HouseCat6D | 5°2cm: 21.3 -> 24.8 | Helped |
| add GCE-Pose global context priors and feature fusion | AG-Pose (DINO) · NOCS-REAL275 | 5°5cm: 64.6 -> 65.1 | Helped |
| add instance geometric prior only (ablation) | AG-Pose (DINO) baseline · HouseCat6D | 5°2cm: 21.3 -> 22.2 | Helped |
| add categorical semantic prior only (ablation) | AG-Pose (DINO) baseline · HouseCat6D | 5°2cm: 21.3 -> 22.7 | Helped |
| use mean-shape geometry with categorical semantics instead of instance-specific geometry (ablation) | GCE-Pose · HouseCat6D | 5°2cm: 24.8 -> 23.4 | Hurt |
| combine instance-specific geometry and categorical semantics (ablation) | AG-Pose (DINO) baseline · HouseCat6D | 5°2cm: 21.3 -> 24.8 | Helped |
| use value features instead of key features for GCE fusion (ablation) | GCE-Pose · HouseCat6D | 5°2cm: 24.8 -> 21.9 | Hurt |
| use value features instead of key features for GCE fusion (ablation) | GCE-Pose · HouseCat6D | 5°5cm: 25.7 -> 23.2 | Hurt |
Surprising findings
- Using key tokens rather than value tokens for the global semantic prior in cross-attention gave better pose accuracy (24.8 vs 21.9 on 5°2cm).
Limitations
- The semantic prior is limited to specific categories and does not scale to instance-level alignment.
- Requires per-category shape prior training and multi-view virtual-camera DINOv2 prototype construction.
Experimental setup
Datasets: HouseCat6D NOCS-REAL275
Benchmarks: HouseCat6D NOCS-REAL275
Metrics: 5°2cm 5°5cm 10°2cm 10°5cm IoU50 IoU75
Baselines: VI-Net SecondPose AG-Pose AG-Pose (DINO) RBP-Pose DPDN Query6DoF GS-Pose
Builds on: AG-Pose deep linear shape model DINOv2 PointNet++ SGPA
Closest papers by meaning
- SchemaPose: RGB-Based Category-Level Pose Estimation with Parametric Category Schema (NeurIPS 2026 · similarity 0.87)
- SecondPose: SE(3)-Consistent Dual-Stream Feature Fusion for Category-Level Pose Estimation (CVPR 2024 · similarity 0.87)
- GeoReF: Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement (CVPR 2024 · similarity 0.87)
- Leveraging Global Stereo Consistency for Category-Level Shape and 6D Pose Estimation from Stereo Images (CVPR 2025 · similarity 0.86)
- GIVEPose: Gradual Intra-class Variation Elimination for RGB-based Category-Level Object Pose Estimation (CVPR 2025 · similarity 0.86)
- 3D-Aware Hypothesis & Verification for Generalizable Relative Object Pose Estimation (ICLR 2024 · similarity 0.86)
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.