Comparative Analysis of Deep Learning Models for Car Part Image Segmentation

Accurate and efficient car part instance segmentation is a fundamental requirement in the automotive industry, with applications ranging from vehicle diagnostics and maintenance to insurance claim assessments. In this study, we present a quantitative approach to car part segmentation, by evaluating and comparing the power of YOLOv8, the Detectron2 Mask R-CNN with Resnet 101 and Mask R-CNN with ResNeXt 101 32×8d configuration+FPN Backbone architecture. Notably, our study focuses on a dataset comprising of 18 distinct car part labels, adding complexity and relevance to the real-world scenarios. Car damage assessment is often accompanied with a multifaceted challenge, with varying damage types and degrees of severity. Identifying and delineating these parts accurately is essential for decision-making in the repair and insurance sectors. The presence of noise factors, such as dirt, grease, and varying lighting conditions, further exacerbates the instance segmentation task. In this study, we have trained and rigorously evaluated our models on a diverse internally labelled dataset consisting of 18 unique car part labels. Our results demonstrate the efficacy of our approach in achieving precise car part segmentation. The Detectron2 Mask R-CNN R101+FPN and ResNeXt 101 32×8d configuration model excelled in real-time car part detection and segmentation, with its powerful backbone architecture exhibited superior performance in handling intricate part boundaries and fine-grained segmentation. Whereas, The YOLO V8 model performed really well in real-time car part detection and segmentation, displaying its versatility in identifying and delineating car parts tasks.