Learning to Match Anchors for Visual Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Current Convolutional Neural Network (CNN)-based object detectors rely on Intersection-over-Union (IoU) thresholds for matching ground-truth objects with anchors.
• This IoU restriction limits flexibility and can impact detection performance.

Purpose of the Study:

• To propose a novel Learning-to-Match (LTM) method to overcome the limitations of IoU-based anchor assignment in object detection.
• To enable a more flexible and effective matching mechanism between objects and anchors.

Main Methods:

• Formulated object detector training within the Maximum Likelihood Estimation (MLE) framework to enable "free" anchor matching.
• Converted detection likelihood into plug-and-play anchor matching loss functions for training.
• Demonstrated the applicability of LTM to both anchor-based and anchor-free object detection architectures.

Main Results:

• LTM detectors consistently outperformed existing methods on the MS COCO dataset with significant accuracy improvements.
• The method showed general applicability across different object detection frameworks.
• Achieved superior performance without introducing additional architectural complexity or parameters.

Conclusions:

• The proposed LTM method effectively breaks the IoU restriction in object detection, leading to enhanced performance.
• LTM offers a computationally efficient and flexible approach to object-feature matching for visual object detection.
• The learnable matching mechanism provides a robust solution for improving CNN-based object detectors.