Research map / 3D vision / Human pose estimation
F2G-Pose: Geometry-Aware Foundation Feature Lifting for Direct RGB-D Category-Level Object Pose Estimation
Jinyu Zhang, Haitao Lin, Jiashu Hou, Xiangyang Xue, Yanwei Fu
Read the paperCheck your idea against 137,217 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 F2G-Pose, a geometry-aware method that lifts foundation-model features into 3D for direct RGB-D category-level object pose estimation.
Problem
Category-level RGB-D object pose estimation (recovering 3D rotation and translation without instance CAD models or reference views) is ambiguous due to intra-class variation, symmetry, and partial observations; deterministic methods suffer error propagation from intermediate structures, while other
Approach
Lifts features from a foundation model into geometry-aware representations for direct pose regression from RGB-D input, avoiding error-prone intermediate structures; only the abstract is available, so specifics are unknown.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add lifted RADIO RGB foundation features (ablation) | F2G-Pose · ROPE | AUC IoU50: 16.2 -> 27.8 | Improved |
| Add MoE Transformer instead of dense FFN (ablation) | F2G-Pose · ROPE | AUC IoU50: 24.0 -> 27.8 | Improved |
| Add camera-frame shape supervision (ablation) | F2G-Pose · ROPE | AUC IoU50: 23.2 -> 27.8 | Improved |
| Use DGCNN tokens instead of global PointNet pooling (ablation) | F2G-Pose · ROPE | AUC IoU50: 17.8 -> 21.6 | Improved |
| Add Transformer context to DGCNN tokens (ablation) | F2G-Pose · ROPE | AUC IoU50: 21.6 -> 24.0 | Improved |
| Swap RADIO backbone for DINOv2 (ablation) | F2G-Pose · ROPE | AUC IoU50: 27.8 -> 27.0 | No clear effect |
| Swap RADIO backbone for CLIP (ablation) | F2G-Pose · ROPE | AUC IoU50: 27.8 -> 26.3 | No clear effect |
| Use top-2 MoE instead of top-1 (ablation) | F2G-Pose · ROPE | AUC IoU50: 25.6 -> 27.8 | Improved |
| Use top-4 MoE instead of top-2 (ablation) | F2G-Pose · ROPE | AUC IoU50: 27.8 -> 28.0 | No clear effect |
| Add load-balancing loss to top-2 MoE (ablation) | F2G-Pose · ROPE | AUC IoU50: 26.0 -> 27.8 | Improved |
| Replace GenPose++ sampling with direct single-pass prediction | F2G-Pose · SOPE | AUC IoU50: 31.9 -> 43.7 | Improved |
| Replace GenPose++ sampling with direct single-pass prediction, speed | F2G-Pose · ROPE | FPS: 4 -> 28 | Improved |
Closest papers by meaning
- SchemaPose: RGB-Based Category-Level Pose Estimation with Parametric Category Schema (NeurIPS 2026 · similarity 0.89)
- SecondPose: SE(3)-Consistent Dual-Stream Feature Fusion for Category-Level Pose Estimation (CVPR 2024 · similarity 0.88)
- GIVEPose: Gradual Intra-class Variation Elimination for RGB-based Category-Level 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)
- Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos (CVPR 2024 · similarity 0.88)
- R^2-Art: Category-Level Articulation Pose Estimation from Single RGB Image via Cascade Render Strategy (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.
