GSDet: Object Detection in Aerial Images Based on Scale Reasoning

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Object detection in aerial images is challenging due to variations in scale, style, and capture scenes.
• Existing methods do not utilize ground sample distance (GSD), a crucial factor for understanding object scale in aerial imagery.

Purpose of the Study:

• To propose the first object detection network that incorporates GSD into the modeling process.
• To enhance object detection performance in aerial images by leveraging GSD and scene context.

Main Methods:

• A two-stage detection framework is enhanced with a GSD identification subnet for probability estimation.
• GSD information is combined with Region of Interest (RoI) sizes to determine physical object dimensions.
• Scene information is integrated to improve inter-category discriminability and adapt inference.

Main Results:

• The proposed method effectively combines GSD and RoI size to estimate physical object size, providing a powerful prior for detection.
• RoI-wise enhanced features are generated by reweighting classification layer weights based on estimated physical size.
• The inclusion of scene information further boosts discriminability and adaptivity in the detection process.

Conclusions:

• The developed object detection network demonstrates significant improvements over existing two-stage methods on the DOTA dataset.
• Incorporating GSD and scene context offers a flexible and effective approach to aerial object detection.
• This work establishes a new baseline for object detection in aerial imagery by utilizing previously overlooked prior knowledge.