Learning And-Or Models to Represent Context and Occlusion for Car Detection and Viewpoint Estimation

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Context and occlusion are significant challenges in car detection and viewpoint estimation.
• Existing methods often struggle to represent complex spatial relationships and occluded object configurations.

Purpose of the Study:

• To develop a novel And-Or model for learning car-to-car context and occlusion.
• To improve the accuracy and robustness of car detection and viewpoint estimation algorithms.

Main Methods:

• Learning an And-Or model with a reconfigurable grammar hierarchy to represent structural and appearance variations.
• A two-stage, weakly supervised learning process involving structure learning (context/occlusion mining, part visibility) and parameter training (Weak-Label Structural SVM).
• Utilizing Dynamic Programming for And-Or model inference.

Main Results:

• The proposed And-Or model demonstrated significant improvements in car detection across four diverse datasets (KITTI, PASCAL VOC2007, Street-Parking, Parking-Lot).
• Achieved comparable performance to state-of-the-art methods in car viewpoint estimation on three datasets (PASCAL VOC2006, 3D car, PASCAL3D+).

Conclusions:

• The And-Or model effectively captures multi-level car context and occlusion configurations.
• This approach offers a powerful framework for enhancing performance in challenging object detection and estimation tasks.