Gliding Vertex on the Horizontal Bounding Box for Multi-Oriented Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object detection commonly uses horizontal bounding boxes, which are inadequate for oriented objects like those in aerial imagery and scene text.
• Existing methods struggle with accurately representing and detecting objects with arbitrary orientations.

Purpose of the Study:

• To propose a simple and effective framework for detecting multi-oriented objects.
• To enhance the accuracy and robustness of object detection for non-horizontal objects.

Main Methods:

• A novel approach is presented that regresses four length ratios to describe the gliding offset of horizontal bounding box vertices.
• An obliquity factor is introduced to address confusion issues with nearly horizontal objects, guiding the choice between horizontal and oriented detection.
• These five extra target variables are integrated into the Faster R-CNN regression head with minimal computational overhead.

Main Results:

• The proposed method achieves superior performance on various multi-oriented object detection benchmarks.
• Demonstrated effectiveness in object detection for aerial images, scene text, and pedestrian detection in fisheye images.
• The framework achieves high accuracy without requiring complex additional components or significant computational cost.

Conclusions:

• The proposed framework offers a simple yet effective solution for multi-oriented object detection.
• The method significantly improves detection accuracy for challenging datasets with oriented objects.
• This approach provides a valuable advancement for computer vision tasks involving non-axis-aligned objects.