Recent Advances in Object Detection, Segmentation, Integration and Relationship Detection with Special Reference to Enhanced Scene Understanding

Abstract

Understanding visual scenes comprehensively remains a central challenge in Computer vision and artificial intelligence. The field has witnessed tremendous evolution from traditional feature-based methods to deep learning architectures capable of simultaneous object detection, precise segmentation and complex relationship modeling. This review synthesizes recent developments across these interconnected domains, with particular emphasis on the YOLO family evolution through YOLOv10, transformer-based detection frameworks including DETR and its variants, advanced segmentation models such as SAM and HQ-SAMand scene graph generation techniques. We examine how multi-task learning and multimodal integration strategies are reshaping scene understanding capabilities. Critical analysis of current limitations—including Computational efficiency, domain generalization, data imbalance and interpretability—guides our discussion of emerging research directions. Foundation models, efficient transformers and zero-shot learning represent promising avenues for advancing robust, scalable scene understanding systems applicable to autonomous vehicles, medical imaging, robotics and intelligent surveillance.