Scaled-YOLOv4: Scaling Cross Stage Partial Network
Chien-Yao Wang Institute of Information Science Academia Sinica, Taiwan Alexey Bochkovskiy Hong-Yuan Mark Liao Institute of Information Science Academia Sinica, Taiwan
Abstract
We show that the YOLOv4 object detection neural network based on the CSP approach, scales both up and down and is applicable to small and large networks while maintaining optimal speed and accuracy. We propose a network scaling approach that modifies not only the depth, width, resolution, but also structure of the network. YOLOv4-large model achieves state-of-the-art results: 55.5% AP (73.4% AP50) for the MS COCO dataset at a speed of $\sim$ 16 FPS on Tesla V100, while with the test time augmentation, YOLOv4-large achieves 56.0% AP (73.3 AP50). To the best of our knowledge, this is currently the highest accuracy on the COCO dataset among any published work. The YOLOv4-tiny model achieves 22.0% AP (42.0% AP50) at a speed of $\sim$ 443 FPS on RTX 2080Ti, while by using TensorRT, batch size = 4 and FP16-precision the YOLOv4-tiny achieves 1774 FPS.
中文速览
用于不同设备的目标检测（object detection）模型往往需要在速度和精度之间做艰难取舍，现有的缩放方法只调整网络深度、宽度和输入分辨率，忽略了网络结构本身的变化。本文提出 Scaled-YOLOv4，在 YOLOv4 基础上引入了一套同时修改深度、宽度、分辨率**以及网络结构**的复合缩放方法，并分别针对低端设备和高端 GPU 设计了轻量版 YOLOv4-tiny 与大模型 YOLOv4-large。实验结果显示，YOLOv4-tiny 在 RTX 2080Ti 上可跑到 443 FPS（配合 TensorRT 甚至达 1774 FPS），而 YOLOv4-large 在 MS COCO 数据集上以约 16 FPS 的速度取得 55.5% AP，刷新了当时已发表工作的最高精度记录。这项工作证明了一套统一的缩放框架能够覆盖从嵌入式芯片到云端 GPU 的全场景需求，为实际部署目标检测模型提供了系统性的设计思路。
原文 arXiv:2011.08036;中英对照 + 大白话阅读 https://aha.fim.ai/paper/2011.08036v2