This article is the continued expansion of the original team of YOLOv4 (including CSPNet and YOLOv4 AB) in the YOLO series. Starting from several different factors that affect model expansion, two types of YOLO suitable for low-end GPU and high-end GPU are proposed. . The YOLO-large proposed in this article has achieved unprecedented accuracy at MSCOCO (the best of the published research results) and can maintain real-time reasoning; the proposed YOLO-tiny is combined with TensorRT+FP16 and other technologies on the RTX 2080Ti graphics card. Achieved an astonishing 1774FPS@batch=4.
Paper title: Scaled-YOLOv4: Scaling Cross Stage Partial Network
Link: https://arxiv.org/2011.08036
Code: https://github.com/WongKinYiu/ScaledYOLOv4
YOLOv4-CSP: https://github.com/WongKinYiu/ ScaledYOLOv4/tree/yolov4-csp
YOLOv4-tiny: https://github.com/WongKinYiu/ScaledYOLOv4/tree/yolov4-tiny
YOLOv4-large: https://github.com/WongKinYiu/ScaledYOLOv4/tree/yolov4-large
Abstract
This article proposes a "Network Scaling" method, which not only adjusts the depth, width, and resolution, but also adjusts the network results. The author calls this method Scaled-YOLOv4.
The resulting YOLOv4-Large achieved SOTA results: 55.4% AP (73.3% AP50) was obtained on the MS-COCO data set, and the inference speed was 15fps@Tesla V100; after adding TTA, the model reached 55.8% AP (73.2% AP50). Up to now, among all public papers, YOLOv-Large has achieved the best index on the COCO data set. The resulting YOLOv4-tiny achieved 22.0% AP (42.0% AP50), and the inference speed was 443fps@TRX 2080Ti; through TensorRT acceleration and FP16 inference, the inference speed can reach 1774fps when batchsize=4.
The main contributions of this paper include the following points:
Design a powerful "network extension" method to improve the performance of small models, which can balance computational complexity and memory usage;
design a simple and effective strategy To expand the large target detector;
analyze the correlation between the model expansion factors and expand the model based on the optimal partition;
experimentally verify: FPN structure is inherently a once-for-all structure
based on the aforementioned analysis and design two efficient models : YOLOv4-tiny and YOLOv4-Large.
