DBF: Dynamic Belief Fusion for Combining Multiple Object Detectors.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object detection systems often use multiple detectors to improve performance.
• Integrating outputs from diverse detectors presents challenges in managing uncertainty and ambiguity.

Purpose of the Study:

• To introduce dynamic belief fusion (DBF), a novel score-level fusion technique for object detection.
• To enhance detection accuracy by effectively integrating inference scores from multiple object detection methods.

Main Methods:

• DBF estimates the ambiguity of each detection score using a confidence model based on precision-recall relationships.
• It calculates probabilities for target, non-target, and intermediate states (basic probability assignment) for each detector's output.
• Dempster's combination rule is employed to optimally fuse these probability distributions.

Main Results:

• Experiments on ARL, PASCAL VOC 07, and 12 datasets demonstrated superior performance of DBF.
• DBF significantly outperformed individual detectors and other baseline fusion approaches in detection accuracy.

Conclusions:

• Dynamic belief fusion (DBF) offers a practical and effective approach for score-level fusion in object detection.
• The method's ability to handle ambiguity and optimally combine detector outputs leads to substantial accuracy improvements.