Combining Faster R-CNN and Model-Driven Clustering for Elongated Object Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Generic object detectors, including R-CNN variants, exhibit reduced performance on objects with low object-region-percentages (ORPs).
• Elongated objects often present challenges due to their low ORPs, leading to lower detection accuracy compared to the average.
• Existing methods struggle with accurately detecting and localizing elongated objects, necessitating specialized approaches.

Purpose of the Study:

• To develop a novel hybrid approach for accurate detection and localization of elongated objects, specifically addressing the challenge of low object-region-percentages (ORPs).
• To improve the performance of object detection systems for elongated shapes, which are often missed or poorly localized by standard algorithms.

Main Methods:

• A hybrid approach combining Faster R-CNN for robust partial object detection and a novel model-driven clustering algorithm for grouping detections and suppressing false positives.
• Training Faster R-CNN with partial region proposals exhibiting suitable and stable ORPs.
• Implementing a deep convolutional neural network (DCNN) for orientation classification of partial detections.
• Utilizing an adaptive model-driven clustering algorithm that initializes an elongated object model from local partial detections and refines it iteratively.

Main Results:

• The proposed method successfully generates tight oriented bounding boxes for elongated objects.
• Experimental evaluation on the COCO dataset and other elongated objects (rigid and non-rigid) demonstrates significant improvements in detection and localization accuracy compared to state-of-the-art methods.
• The hybrid approach effectively addresses the low ORP problem for elongated object detection.

Conclusions:

• The hybrid Faster R-CNN and model-driven clustering approach provides a robust solution for detecting and localizing elongated objects with low ORPs.
• This method offers a substantial advancement in computer vision for handling challenging object shapes, outperforming existing techniques.
• The developed technique enhances the reliability and precision of object detection systems for a wider range of object types and aspect ratios.