YOLOv7: Trainable Bag-of-Freebies Sets New State-of-the-Art for Real-Time Object Detectors

The paper addresses real-time object detection, one of the most important research topics in computer vision. Observing two research topics that spawned from recent state-of-the-art work on architecture optimization and training optimization, the authors propose a trainable bag-of-freebies oriented solution: flexible and efficient training tools are combined with a proposed architecture and a compound scaling method to build YOLOv7.

YOLOv7 surpasses all known object detectors in both speed and accuracy across the range from 5 FPS to 120 FPS, and achieves 56.8% AP, the highest accuracy among all known real-time object detectors running at 30 FPS or higher on a V100 GPU. This established a new state of the art for real-time detection, and the source code was released publicly on GitHub.