YOLOv4

YOLOv4 is an open-source real-time object detection model released in 2020 as part of the You Only Look Once (YOLO) family. It was developed by Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Chen, with the goal of delivering higher accuracy without sacrificing speed on common hardware. The model is designed for practical applications where fast inference is essential, such as video surveillance and autonomous systems.

The architecture combines a robust backbone, a feature fusion neck, and efficient detection heads. The backbone is CSPDarknet53, which uses Cross-Stage Partial connections to improve gradient flow and efficiency. The neck employs a path-aggregation network (PANet) to fuse features from multiple scales, while a spatial pyramid pooling (SPP) layer enhances multi-scale context. The overall design emphasizes a balance between detection quality and real-time performance.

Training and data augmentation are central to YOLOv4, organized around the concepts of Bag of Freebies and Bag of Specialties. Freebies include training-time improvements such as mosaic data augmentation, multi-scale training, and regularization techniques that do not add inference cost. Specialties cover components added during inference, such as certain architectural blocks, which can improve accuracy with modest computational impact. The approach aims to maximize performance on standard datasets like COCO without requiring specialized hardware.

In evaluation, YOLOv4 demonstrated improved accuracy over its predecessors while maintaining real-time speeds on conventional GPUs, contributing to its broad adoption in research and industry. It served as a practical benchmark for real-time object detection and influenced subsequent developments in the YOLO series and related detection frameworks, highlighting a path toward higher accuracy at feasible inference speeds.