Keypoint-Based Disentangled Pose Network for Category-Level 6-D Object Pose Tracking.

Area of Science:

• Computer Vision
• Robotics
• 3-D Reconstruction

Background:

• Category-level 6-D object pose tracking is a complex problem in 3-D computer vision.
• Keypoint-based methods are effective but often treat rotation and translation estimation uniformly.
• Existing approaches use neural networks for keypoint estimation followed by least-squares optimization for pose change.

Purpose of the Study:

• To propose a novel keypoint-based disentangled pose network for 6-D object pose tracking.
• To improve accuracy by separately estimating 3-D rotation and 3-D translation.
• To address limitations of current methods that do not differentiate between rotation and translation estimation.

Main Methods:

• A keypoint-based disentangled pose network is introduced.
• The network directly estimates 3-D translation.
• Singular value decomposition is used to indirectly calculate 3-D rotation based on estimated keypoints.

Main Results:

• The proposed method demonstrates superior performance in 6-D object pose tracking.
• Experiments were conducted on the NOCS-REAL275 dataset.
• The disentangled approach shows significant improvements over existing methods.

Conclusions:

• The keypoint-based disentangled pose network effectively addresses challenges in 6-D object pose tracking.
• Separating rotation and translation estimation leads to enhanced accuracy.
• The method offers a promising advancement for 3-D computer vision applications.