Scaling 3D Compositional Models for Robust Classification and Pose Estimation.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Object Recognition

Background:

• Deep learning models struggle with out-of-distribution data (e.g., weather, occlusion) in object classification and 3D pose estimation.
• Neural Mesh Models show promise for robustness but face scalability issues with numerous object classes due to quadratic training complexity.

Purpose of the Study:

• To develop a scalable and robust 3D compositional model for object classification and pose estimation.
• To address the quadratic scaling problem in training Neural Mesh Models for large numbers of object classes.

Main Methods:

• Restructured per-vertex contrastive learning into within-class and between-class comparisons.
• Introduced dynamic decoupling of between-class contrasts, enhancing focus on confused classes.
• Leveraged object compositionality to reduce training time and improve performance.

Main Results:

• Achieved state-of-the-art performance in simultaneous classification and pose estimation.
• Demonstrated superior robustness to out-of-distribution testing (occlusion, weather, synthetic data).
• Showcased effective generalization to previously unseen object classes.

Conclusions:

• The proposed large-scale 3D compositional model offers significant improvements in performance and robustness over existing methods.
• The strategy effectively scales Neural Mesh Models to hundreds of object classes.
• This approach advances reliable 3D object recognition in challenging, real-world conditions.