YOLO*C — Adding context improves YOLO performance

Abstract

You Only Look Once (YOLO) algorithms deliver state-of-the-art performance in object detection. This research proposes a novel one-stage YOLO-based algorithm that explicitly models the spatial context inherent in traffic scenes. The new YOLO*C algorithm introduces the MCTX context module and integrates loss function changes, effectively leveraging rich global context information. The performance of YOLO*C models is tested on BDD100K traffic data with multiple context variables. The results show that including context improves YOLO detection results without losing efficiency. Smaller models report the most significant improvements. The smallest model accomplished more than a 10% increase in mAP .5 compared to the baseline YOLO model. Modified YOLOv7 outperformed all models on mAP .5, including two-stage and transformer-based detectors, available at the dataset zoo. The analysis shows that improvement mainly results from better detection of smaller traffic objects, which presents a significant detection challenge within the complex traffic environment.