SCRDet++: Detecting Small, Cluttered and Rotated Objects via Instance-Level Feature Denoising and Rotation Loss...

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Object detection is challenged by small, cluttered, and rotated objects.
• Traditional methods struggle with accurate bounding boxes for rotated objects, leading to background contamination.

Purpose of the Study:

• To enhance object detection for small, cluttered, and rotated objects.
• To introduce instance-level denoising and address rotation variations in object detection.

Main Methods:

• Instance-level denoising on feature maps to improve small object detection.
• Incorporating a novel IoU constant factor into smooth L1 loss to resolve boundary issues caused by angular periodicity and edge exchangeability.
• Developing the SCRDet++ detector by combining denoising and rotation-handling techniques.

Main Results:

• The proposed SCRDet++ detector demonstrates effectiveness on various datasets, including aerial, natural, scene text, and traffic light images.
• Experiments were conducted on DOTA, DIOR, UCAS-AOD, COCO, ICDAR2015, BSTLD, and the newly released S²TLD datasets.
• The new S²TLD dataset, featuring 5,786 images and 14,130 traffic light instances, was made publicly available.

Conclusions:

• The integration of denoising and a novel IoU factor significantly enhances object detection performance.
• SCRDet++ effectively addresses the challenges posed by small, cluttered, and rotated objects.
• The study contributes a new dataset and an improved detection method for computer vision applications.