Very Large Scale Cross Corpus Object Detection Applications with the Latest Yolo Models on Single Object Lasot Dataset

Abstract

In this study, we explore the latest YOLO models, namely, YOLOv8X, YOLOv9e, YOLOv10X, YOLO11X, and YOLO12X in the single object detection task with LaSOT dataset. YOLO is a great breakthrough in object detection and fills in the gap between high performance and accuracy in real time applications. YOLO tries to maintain accuracy with a reasonable speed which makes it very suitable for real time object recognition applications. This study evaluates the latest YOLO models in terms of accuracy and speed in a very large-scale object detection task on the single object LaSOT dataset. All YOLO models used in this study are pretrained on the COCO dataset and evaluated on the LaSOT dataset. COCO and LaSOT datasets have 30 common classes which include more than 1.5 million image samples. Cross corpus experiments are very formidable challenges and ultimate test for the generalizability and integrity of machine learning models where the models are trained in a dataset and tested in another dataset. Results of the experiments show that YOLOv9e is the best model in terms of accuracy metrics by 0.3859 mAP@0.5 and 0.6490 recall, however, YOLOv10X is the fastest YOLO model with 201.18 fps.