Omni6DPose: A Comprehensive Benchmark and Model for 6D Object Pose Estimation and Tracking

Introduction

6D object pose estimation, a critical task in computer vision, has long been constrained by the lack of large-scale, diverse datasets. This scarcity hampers model evaluation and limits research progress across various domains. To address these issues, researchers from Peking University and other institutions have introduced Omni6DPose, a comprehensive dataset and model designed to advance 6D object pose estimation and tracking.

Dataset Overview

Omni6DPose is divided into three main components:

• ROPE (Real 6D Object Pose Estimation Dataset): Contains 332K images annotated with over 1.5M annotations across 581 instances in 149 categories.
• SOPE (Simulated 6D Object Pose Estimation Dataset): Comprises 475K images created in a mixed reality setting with depth simulation, annotated with over 5M annotations across 4162 instances in the same 149 categories.
• Manually Aligned Real Scanned Objects: Used in both ROPE and SOPE to ensure consistency.

The dataset's diversity in object categories, large scale, and variety in object materials make it a significant resource for researchers. The substantial variations and ambiguities within the dataset present unique challenges that are addressed by the accompanying model, GenPose++.

Model: GenPose++

GenPose++ is an enhanced version of the state-of-the-art category-level pose estimation framework. It incorporates two pivotal improvements:

• Semantic-aware Feature Extraction: This technique leverages semantic information to improve feature representation, making it more robust to variations in object appearance and environment.
• Clustering-based Aggregation: This method groups similar features to refine pose estimates, reducing ambiguity and improving accuracy.

Benchmarking Analysis

The paper provides a comprehensive benchmarking analysis to evaluate the performance of previous methods on the large-scale Omni6DPose dataset. The analysis covers:

• 6D Object Pose Estimation: Evaluates how well models can estimate the 6D pose (position and orientation) of objects.
• Pose Tracking: Assesses the ability of models to track object poses over time, which is crucial for applications like augmented reality and robotics.

The results highlight the strengths and limitations of existing methods, providing valuable insights for future research.

Key Contributions

• Large-scale Dataset: Omni6DPose offers a vast collection of annotated images across diverse object categories and materials.
• Enhanced Model: GenPose++ introduces semantic-aware feature extraction and clustering-based aggregation to improve pose estimation accuracy.
• Comprehensive Benchmarking: The paper provides detailed performance evaluations, facilitating the development and comparison of new methods.

Conclusion

Omni6DPose represents a significant step forward in 6D object pose estimation and tracking. By providing a large-scale, diverse dataset and an enhanced model, it addresses key challenges in the field and paves the way for more robust and versatile computer vision applications.