Detection and localization of multi-scale and oriented objects using an enhanced feature refinement algorithm.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Object detection is crucial in computer vision.
• Existing rotation detectors face challenges with arbitrary orientations, dense arrangements, and loss discontinuity.

Purpose of the Study:

• To present a novel single-stage rotation detector for accurate detection of oriented and multi-scale objects.
• To address limitations of current rotation detectors in cluttered and diverse scenarios.

Main Methods:

• A progressive regression approach using both horizontal and rotating anchors.
• Integration of a feature refinement module to enhance feature angulation and reduce bounding boxes.
• A novel adjustable loss function to mitigate loss discontinuity issues.

Main Results:

• The proposed detector achieves outstanding performance on benchmark datasets.
• Demonstrates significant improvements in both speed and accuracy compared to state-of-the-art methods.
• Effectively handles arbitrarily oriented objects and densely packed scenarios.

Conclusions:

• The novel single-stage rotation detector offers superior performance for object detection tasks.
• The developed methods, including the adjustable loss function, are extendable to other detector types.
• This work advances the field of computer vision for object detection applications.