Benchmarking YOLO Variants on an Unseen Video: A Comparison of Inference Speed, GFLOPs, and Recall

Object detection serves as a fundamental task in the field of computer vision and recent developments in YOLO family aim to enhance the real time detection performance. However, the generalization performance of recent YOLO models on unseen video datasets remains underexplored. Analysing the model performance on unseen datasets is essential for assessing robustness in real world deployments. This work provides a systematic comparison of recent YOLO versions and their variants. The study evaluates YOLOv9, YOLOv10 and YOLOv11 models on an unseen MOT20 video dataset for multi pedestrian detection. Pedestrian detection is important because it forms the basis for many computer vision tasks involving human interaction, crowd monitoring, behaviour analysis, and traffic management the models and their variants are evaluated using the metrics: recall, inference speed and GFLOPs. The experimental results indicates that the variant YOLOv9-m achieves highest recall of 43.9% among all evaluated models, while YOLOv11-n showed marginally lower recall value of 40.9%. However, YOLOv11-n exhibits significantly faster inference speed (9.6ms per image) compared to YOLOv9-m (26.3ms per image) and fewer computational resources-(6.5 Vs 131.3). In contrast YOLOv10 exhibits significantly lower recall (28%) despite its increased efficiency. These findings highlight the inherent trade-offs between accuracy-efficiency in recent YOLO architectures. The study offers the understanding of strengths and limitations of modern YOLO models, aiding in model selection for real time computer vision applications.