Corr-Track: Category-Level 6D Pose Tracking with Soft-Correspondence Matrix Estimation.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Category-level pose tracking enables object pose estimation without instance-specific models, crucial for augmented reality (AR) and virtual reality (VR).
• A key challenge is achieving robust generalization for neural networks to accurately predict poses of previously unseen objects within a category.

Purpose of the Study:

• To introduce Corr-Track, a novel category-level 6D pose tracking method designed for accurate object tracking from depth video streams.
• To address the generalization challenge in category-level pose tracking by developing a robust neural network training approach.

Main Methods:

• Utilizes direct soft correspondence constraints to train a neural network for estimating bidirectional soft correspondences between sparse point clouds from different frames.
• Introduces a soft correspondence matrix and employs direct spatial point-to-point correspondence representations.
• Proposes a 'point cloud expansion' strategy to mitigate the 'point cloud shrinkage' issue inherent in soft correspondences, ensuring accurate shape reproduction.

Main Results:

• Corr-Track achieved superior performance on the NOCS-REAL275 and Wild6D datasets compared to existing methods.
• Cross-category experiments demonstrated the method's strong generalization capability, indicating its potential for diverse applications.

Conclusions:

• Corr-Track offers an effective solution for category-level 6D pose tracking using depth data.
• The proposed soft correspondence and point cloud expansion strategies enhance tracking accuracy and generalization performance for unseen objects.