Research map / 3D vision / Human pose estimation

Emergence of a Shared Canonical Object Frame from In-the-Wild Videos

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.

Shows that a shared canonical object frame can emerge from self-supervised dense correspondence learning on ~160k in-the-wild UCO3D object videos, routed through a coarse canonical cube mesh with per-sequence SfM alignment, without canonical pose labels or category conditioning.

Problem

Category-level object pose estimation requires a shared canonical frame across instances, which traditionally needs manual or synthetic canonical pose supervision that does not scale in category and instance diversity.

Approach

Route all training sequences through a shared coarse canonical mesh (cube) that carries no category detail; learn pixel-to-mesh-vertex correspondences and estimate per-sequence SfM-to-canonical rotations via inverse-entropy-weighted Wahba/SVD alignment, so a common frame emerges from multi-view consistency and DINOv3 semantic features.

Key results

• Achieves best macro-average Acc@30 of 61.8% and median error 25.5 deg across five benchmarks, versus 58.6% for OrientAnything V2 (trained with canonical labels).
• On Pascal3D+ and ObjectNet3D category-specific comparison, reaches 87.6% and 70.0% Acc@30 (mean 78.5%), above Common3D's 66.1% mean.
• Ablation: removing learned alignment drops Pascal3D+ Acc@30 from 79.8% to 55.0%; using a single view per sequence collapses it to 28.0%.

Surprising findings

• PCA-only initialization without learned alignment improves on continuously symmetric objects (59.1% to 65.0%) while degrading non-symmetric ones badly.
• Method trails canonically supervised OrientAnything V2 on Omni6DPose, attributed to the large share of symmetric objects.

Limitations

• Symmetric objects perform worse because pseudo-labels are noisy when orientation is ambiguous.
• Translation is recovered only up to a global scale; no metric scale prediction.
• Requires per-category convention mapping estimated on held-out data for fair evaluation.

Experimental setup

Datasets: UCO3D Pascal3D+ ObjectNet3D ImageNet3D

Benchmarks: REAL275 Omni6DPose Objectron Pascal3D+ ImageNet3D ObjectNet3D

Metrics: median rotation error Acc@30

Baselines: OrientAnything V1 OrientAnything V2 QWEN3-VL ZSP UOP3D Common3D

Builds on: DINOv3 UCO3D Common3D UOP3D EPnP LoRA

Closest papers by meaning

• Self-Supervised Geometric Correspondence for Category-Level 6D Object Pose Estimation in the Wild (ICLR 2023 · similarity 0.90)
• Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos (CVPR 2024 · similarity 0.90)
• One-shot 3D Object Canonicalization based on Geometric and Semantic Consistency (CVPR 2025 · similarity 0.88)
• TAGA: Self-supervised Learning for Template-free Animatable Gaussian Articulated Model (CVPR 2025 · similarity 0.86)
• Articulation in Prime: Primitive-Based Articulated Object Understanding from a Single Casual Video (NeurIPS 2026 · similarity 0.85)
• Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency (CVPR 2025 · similarity 0.85)