Learning Boundary Continuity-Aware Gaussian Encoder for Oriented Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Oriented object detection is vital for rotation-sensitive tasks.
• Existing methods suffer from boundary discontinuity issues due to angle vector outputs.
• This discontinuity can magnify differences between visually similar bounding boxes.

Purpose of the Study:

• To propose a Boundary Continuity-aware Gaussian Encoder (BCGE) for oriented object detection.
• To address and overcome the boundary discontinuity issue in current detection methods.
• To enhance the accuracy and robustness of oriented bounding box prediction.

Main Methods:

• BCGE directly predicts target Gaussian distributions for object proposals.
• It learns oriented bounding boxes as integrated 2-D matrices, ensuring boundary continuity.
• A transformation from Gaussian representation back to boxes is proposed and extended to the complex domain for neural network adaptation.

Main Results:

• BCGE effectively resolves the boundary discontinuity problem in oriented object detection.
• The method demonstrates consistent effectiveness across five popular datasets: DOTA, UCAS-AOD, HRSC2016, SSDD, and HRSID.
• BCGE functions as a plug-and-play module, adaptable to various existing oriented detectors.

Conclusions:

• The proposed BCGE significantly improves oriented object detection performance.
• Its Gaussian-based approach offers a more robust and continuous representation of bounding boxes.
• BCGE provides a versatile and effective solution for enhancing rotation-sensitive detection tasks.