YOLOv4: A Breakthrough in Real-Time Object Detection
Athulya Sundaresan Geetha
TL;DR
YOLOv4 addresses the need for accurate real-time object detection by integrating a comprehensive set of improvements across data processing, architectural design, and training strategies. The main approach combines Bag of Freebies (BoF) techniques such as Mosaic augmentation, DropBlock regularization, and CIoU loss with Bag of Specials (BoS) components like SPP, PANet-based fusion, and attention mechanisms within a CSPDarkNet53 backbone. The key contributions include a detailed architecture that balances speed and accuracy, improved multi-scale detection, and robust training methodologies that achieve $AP=43.5\%$ on COCO with $AP_{50}=65.7\%$ at ~65 FPS on high-end GPUs. This work demonstrates strong practical impact for real-time applications in surveillance, autonomous systems, and industrial inspection by delivering efficient, scalable object detection.
Abstract
YOLOv4 achieved the best performance on the COCO dataset by combining advanced techniques for regression (bounding box positioning) and classification (object class identification) using the Darknet framework. To enhance accuracy and adaptability, it employs Cross mini-Batch Normalization, Cross-Stage-Partial-connections, Self-Adversarial-Training, and Weighted-Residual-Connections, as well as CIoU loss, Mosaic data augmentation, and DropBlock regularization. With Mosaic augmentation and multi-resolution training, YOLOv4 achieves superior detection in diverse scenarios, attaining 43.5\% AP (in contrast, 65.7\% AP50) on a Tesla V100 at ~65 frames per second, ensuring efficiency, affordability, and adaptability for real-world environments.