Improving Performance and Adaptivity of Anchor-Based Detector Using Differentiable Anchoring With Efficient Target...

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Anchor-based object detection methods rely on predefined anchor boxes.
• Predefined anchors limit detector performance and adaptability across datasets.
• Existing methods for learning anchor shapes still depend on predefined anchors.

Purpose of the Study:

• To propose a learning anchoring scheme that removes dependency on predefined anchors.
• To improve the performance and adaptability of object detection models.
• To develop a universal object detection framework.

Main Methods:

• Introduced differentiable anchoring, a new learning anchoring scheme.
• Developed a target generation method using L_p norm ball approximation and optimization difficulty-based pyramid level assignment.
• Integrated the scheme into Faster RCNN, RetinaNet, and SSD architectures.

Main Results:

• Achieved significant improvements in mean Average Precision (mAP) on the MS COCO 2017 test-dev set (2.8% for Faster RCNN, 2.1% for RetinaNet, 2.3% for SSD).
• Demonstrated superior performance and adaptability without hyperparameter tuning or specialized optimization.
• Showcased competitive results on challenging tasks like tiny face and text detection.

Conclusions:

• Differentiable anchoring offers a universal solution for object detection, removing predefined anchor dependencies.
• The method enhances detector performance and adaptability across diverse datasets and tasks.
• This approach provides a robust and flexible framework for future object detection research.