Multi-Oriented Object Detection in Aerial Images With Double Horizontal Rectangles.

Area of Science:

• Computer Vision
• Machine Learning
• Object Detection

Background:

• Existing methods for multi-oriented object detection often use quadrilateral or rotated rectangle representations.
• These representations can lead to non-unique object encoding due to vertex ordering, angular periodicity, and edge exchangeability, introducing discontinuity problems.

Purpose of the Study:

• To propose a novel representation for multi-oriented object detection that resolves the discontinuity problem.
• To enhance the accuracy and robustness of object detection across arbitrary orientations.

Main Methods:

• Introduced the Double Horizontal Rectangle (DHRec) representation for encoding multi-oriented objects.
• Defined DHRec using ordered horizontal and vertical coordinates of object vertices.
• Regressed area ratios between regions to guide the decoding of oriented objects from the predicted DHRec.

Main Results:

• The DHRec representation ensures uniqueness, eliminating the discontinuity issue inherent in previous methods.
• The proposed method demonstrated significant improvements over existing baselines.
• Achieved superior performance compared to state-of-the-art object detection methods.

Conclusions:

• The DHRec representation offers a unique and robust solution for detecting multi-oriented objects.
• This approach effectively addresses the discontinuity problem, leading to more accurate object detection.
• The method shows strong potential for advancing the field of oriented object detection.