Diff9D: Diffusion-Based Domain-Generalized Category-Level 9-DoF Object Pose Estimation.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Category-level nine-degrees-of-freedom (9-DoF) object pose and size estimation is vital for augmented reality and robotic manipulation.
• Existing category-level methods generalize to unseen objects but require extensive real-world data collection and labeling.
• Domain generalization remains a significant challenge, limiting the applicability of models trained on specific datasets.

Purpose of the Study:

• To develop a diffusion-based paradigm for domain-generalized category-level 9-DoF object pose estimation.
• To leverage the generalization capabilities of diffusion models to overcome domain shift issues.
• To enable accurate pose estimation in real-world scenarios using models trained solely on synthetic data.

Main Methods:

• Proposed a novel diffusion model for 9-DoF object pose estimation, framing it as a generative task.
• Trained the model exclusively on rendered synthetic data, eliminating the need for real-world labeled data.
• Utilized the Denoising Diffusion Implicit Model (DDIM) for efficient reverse diffusion, achieving near real-time performance in as few as 3 steps.
• The model does not require 3D shape priors during training or inference.

Main Results:

• Achieved state-of-the-art domain generalization performance on two benchmark datasets.
• Demonstrated successful application in a real-world robotic grasping system.
• Near real-time performance was achieved with a 3-step reverse diffusion process.

Conclusions:

• The proposed diffusion-based approach effectively addresses the domain generalization challenge in category-level 9-DoF object pose estimation.
• Training solely on synthetic data enables robust generalization to real-world scenes.
• The method offers a practical solution for applications like robotic manipulation and augmented reality.