Research map / 3D vision / Human pose estimation

GenCOPE: Syn2Real Generalized Category-Level Object Pose Estimation for Robotic Picking

NeurIPS 2026 · 3D vision › Human pose estimation · approach: Category-level object pose estimation · 0 citations (OpenAlex)

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.

Trains a lightweight RGB-D category-level pose estimator solely on synthetic CAMERA25 using LLM-generated 2D/3D semantic descriptions as contrastive supervision, and transfers zero-shot to REAL275, Wild6D and a robotic picking setup.

Problem

Category-level 9DoF object pose estimation normally needs costly real-world annotated training data per novel category; the synthetic-to-real domain gap, mainly texture appearance, hurts generalization.

Approach

Use LLM-generated intra-class part-structured 2D and 3D semantic descriptions, encoded by CLIP text encoder, as a supervised contrastive target for ResNet18 and PointNet global features during training only, followed by a 2D-3D cross-transformer fusion and decoupled pose decoder that uses only global features.

Key results

• On REAL275 (Mask R-CNN trained on ImageNet), GenCOPE reaches 42.6 3D75 and 39.7 at 5deg2cm, outperforming Diff9D (41.7, 35.3) and CleanPose (39.9, 39.5) with CAMERA25-only training.
• On Wild6D, GenCOPE achieves 38.72 3D75 and 42.74 at 10deg5cm, the best among compared methods including Diff9D.
• GenCOPE uses 2.89 GB inference GPU memory and runs at 40.68 FPS, versus 10.04 GB / 24.33 FPS for CleanPose and 9.60 GB / 17.20 FPS for Diff9D.

Surprising findings

• Replacing Gemini-3 with Qwen2.5-VL-7B or GPT-4 for generating semantic descriptions yields comparable accuracy, suggesting robustness to the choice of LLM.
• Text features injected directly into the encoders (residual addition) slightly degrade performance, whereas the semantic consistency loss alone gives gains.

Limitations

• Does not explicitly handle pose equivalence for strongly symmetric categories.
• Failure cases under severe occlusion, transparent objects with incomplete depth, and extremely low light.

Experimental setup

Datasets: CAMERA25 REAL275 Wild6D

Benchmarks: REAL275 Wild6D self-collected robotic picking scenes

Metrics: IoU_3D at 75% 5deg2cm 5deg5cm 10deg2cm 10deg5cm mAP GPU memory (GB) FPS

Baselines: CleanPose Diff9D DPDN UDA-COPE TTA-COPE SAR-Net SecondPose VI-Net SGPA GPV-Pose

Builds on: CLIP ResNet PointNet SPD DiffusionNOCS

Closest papers by meaning

• Vision Foundation Model Enables Generalizable Object Pose Estimation (NeurIPS 2024 · similarity 0.88)
• 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)
• One-shot 3D Object Canonicalization based on Geometric and Semantic Consistency (CVPR 2025 · similarity 0.86)
• SUGAR: Pre-training 3D Visual Representations for Robotics (CVPR 2024 · similarity 0.86)
• OmniManip: Towards General Robotic Manipulation via Object-Centric Interaction Primitives as Spatial Constraints (CVPR 2025 · similarity 0.86)