On Boundary Discontinuity in Angle Regression Based Arbitrary Oriented Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Oriented object detection is crucial for autonomous driving and remote sensing.
• Existing methods struggle with boundary discontinuity due to direct angle regression.

Purpose of the Study:

• To propose a novel angle coder, the phase-shifting coder (PSC), to address boundary discontinuity in oriented object detection.
• To develop a dual-frequency version (PSCD) for improved orientation prediction of diverse object shapes.

Main Methods:

• Introduced the phase-shifting coder (PSC) for continuous and differentiable angle encoding.
• Developed a dual-frequency version (PSCD) to handle rotational symmetry in object boundaries.
• Integrated PSC/PSCD with existing backbone detectors and loss functions (e.g., Gaussian, RotatedIoU).

Main Results:

• PSC effectively eliminates boundary discontinuity in angle prediction.
• PSCD accurately predicts orientations for both elongated and square-like objects.
• Experiments demonstrate performance improvements across various datasets and backbone detectors.

Conclusions:

• The proposed PSC and PSCD offer a boundary-discontinuity-free approach to oriented object detection.
• These methods can significantly boost the performance of existing oriented object detection frameworks.
• The approach shows strong potential for applications demanding high-quality bounding boxes.