RatioNet: Ratio Prediction Network for Object Detection.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Anchor-free object detectors in remote sensing struggle with false positives and imbalanced samples due to single-oriented features and key point-based strategies.
• Existing methods often fail to adequately address these challenges, limiting detection accuracy.

Purpose of the Study:

• To develop a simple, effective, and parameter-efficient anchor-free object detection approach for remote sensing images.
• To mitigate issues of sample imbalance and false boxes inherent in current detectors.

Main Methods:

• RatioNet assigns all points within ground-truth boxes as positive samples to combat sample imbalance.
• A novel regression approach predicts object width, height, and location ratios (l_ratio, t_ratio) using global object features to reduce false boxes.
• The introduction of ratio-center weighting enhances the preservation of high-quality bounding boxes.

Main Results:

• RatioNet demonstrates improved accuracy and reduced false positives compared to existing methods.
• The proposed approach achieves 49.7% Average Precision (AP) on the MS-COCO test-dev dataset.
• The method is effective in handling the complexities of object detection in remote sensing imagery.

Conclusions:

• RatioNet offers a significant advancement in anchor-free object detection for remote sensing.
• The method's efficiency in parameter usage and high accuracy make it a valuable tool for analyzing remote sensing data.
• Future work could explore RatioNet's application in diverse remote sensing scenarios and datasets.