Learning Oriented Object Detection via Naive Geometric Computing.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Accurate detection and rotation estimation of oriented objects are vital for remote sensing image analysis.
• Existing methods often rely on single or limited ground truth (GT) values for supervision, potentially limiting accuracy and robustness.
• Joint supervision using additional constraints for proposal and rotation regression can enhance oriented object detection.

Purpose of the Study:

• To propose a novel mechanism for simultaneous regression of horizontal proposals, oriented proposals, and object rotation angles.
• To enhance proposal quality and overall detection performance through a consistent, geometrically constrained training approach.
• To introduce an oriented center prior guided label assignment strategy for improved proposal generation.

Main Methods:

• A mechanism for simultaneous regression of horizontal and oriented object proposals and their rotation angles using naive geometric computing.
• An oriented center prior guided label assignment strategy to improve the quality of training proposals.
• Joint supervision during training incorporating these geometric constraints.

Main Results:

• The proposed method significantly outperforms the baseline in oriented object detection across six diverse datasets.
• New state-of-the-art results are achieved without increasing computational overhead during inference.
• The approach demonstrates improved accuracy and robustness in detecting oriented objects and their rotations.

Conclusions:

• The proposed joint supervision mechanism provides a simple, intuitive, and effective way to enhance oriented object detection.
• The method offers significant performance gains and establishes new state-of-the-art benchmarks.
• The approach is readily implementable and computationally efficient for practical applications in remote sensing.