COMPARE ONE-STAGE AND TWO-STAGE SEGMENTATION FOR DETECTING VEHICLE DAMAGE IN AN IMAGE

Abstract

The rapid growth of the automotive industry and the increasing demand for efficient accident damage assessment have led to a need for accurate and automated vehicle damage detection systems. This paper presents a comparative analysis of one-stage and two-stage segmentation methods for detecting various types of damage in vehicles using computer vision and deep learning. The main goal is to determine which approach to segmentation brings more value and better results when detecting types of damage such as dents, scratches, cracks, broken glass, broken lights, and tires flat

The Car Damage Detection dataset (CarDD), consisting of 4,000 images and over 9,000 annotated damage instances, was used for training and evaluation. Two models were chosen for comparison: YOLOv8, representing the one-stage segmentation approach, and Mask R-CNN, as the two-stage segmentation method. YOLOv8 offers faster inference and simpler architecture, while Mask R-CNN provides higher segmentation precision due to its refined region proposal mechanism and RoIAlign technique.

Experiments show that Mask R-CNN outperforms YOLOv8 in terms of segmentation accuracy, especially in complex damage scenarios. Metrics such as mean Average Precision and confusion matrices were used to analyze productivity for different types of damage. While YOLOv8 was highly effective in detecting simple types of damage such as dents and scratches, Mask R-CNN consistently performed better in all categories, especially in minimizing background blur.

This study concludes that while YOLOv8 is suitable for real-time applications due to its speed, Mask R-CNN is better suited for scenarios where detection accuracy is critical, such as insurance assessments or automated vehicle inspection systems.