Research map / 3D vision / Human pose estimation

SchemaPose: RGB-Based Category-Level Pose Estimation with Parametric Category Schema

NeurIPS 2026 · 3D vision › Human pose estimation · approach: 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 SchemaPose, an end-to-end DINO-based query detector that infers a parametric category schema vector from RGB and uses it at proposal and decoder stages to guide category-level 9D pose prediction, trained with a schema-driven synthetic dataset.

Problem

RGB-only category-level 9D pose estimation must jointly infer translation, rotation, and size across unseen instances, but existing methods either rely on canonical NOCS geometry or absorb category structure implicitly into features.

Approach

Introduce an explicit parametric category schema (discrete part choices plus continuous geometric parameters) compressed into a compact schema vector that is predicted from encoder and decoder features and fed back as structural guidance to all pose branches in a dual-stage query-based head.

Key results

• On REAL275, SchemaPose reaches 68.4 10°10cm vs. 52.8 for the strongest prior RGB method YOPO.
• On CAMERA25, SchemaPose reaches 43.1 10°10cm vs. 38.7 for YOPO.
• Ablation: removing schema (w/o Schema) drops REAL275 10°10cm from 68.4 to 63.8; shuffled schema drops it to 63.1.

Surprising findings

• A same-dimensional learned latent vector without schema supervision underperforms the true schema, supporting that gains come from schema semantics rather than extra parameters.
• Increasing SchemaData from 135K to 270K images gives no further gain.

Limitations

• Schema covers only six NOCS categories with hand-designed templates, limiting generality to new categories.
• REAL275 schema annotations required manual effort (about 20 minutes per object) for the training split.

Experimental setup

Datasets: CAMERA25 REAL275 SchemaData ShapeNetCore COCO

Benchmarks: CAMERA25 REAL275

Metrics: IoU50 IoU75 10cm 10deg 10deg10cm mAP

Baselines: MSOS OLD-Net DMSR LaPose GIVEPose MonoDiff9D DA-Pose Fischer et al. LanCOPE YOPO

Builds on: DINO NOCS DINO-based pose estimation (Simonelli et al. disentangled monocular pose) 6D rotation representation

Closest papers by meaning

• GIVEPose: Gradual Intra-class Variation Elimination for RGB-based Category-Level Object Pose Estimation (CVPR 2025 · similarity 0.91)
• CAP-Net: A Unified Network for 6D Pose and Size Estimation of Categorical Articulated Parts from a Single RGB-D Image (CVPR 2025 · similarity 0.89)
• R^2-Art: Category-Level Articulation Pose Estimation from Single RGB Image via Cascade Render Strategy (AAAI 2025 · similarity 0.89)
• Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos (CVPR 2024 · similarity 0.89)
• Universal Features Guided Zero-Shot Category-Level Object Pose Estimation (AAAI 2025 · similarity 0.89)
• SecondPose: SE(3)-Consistent Dual-Stream Feature Fusion for Category-Level Pose Estimation (CVPR 2024 · similarity 0.88)