Scaled-YOLOv4: Scaling Cross Stage Partial Network

By Chien-Yao Wang, Alexey Bochkovskiy, Hongpeng Liao. 2020

1,445 citations recorded by the source corpus.

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 ~ 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 ~443 FPS on RTX 2080Ti, while by using TensorRT, batch size = 4 and FP16-precision the YOLOv4-tiny achieves 1774 FPS.

Subfields: Computer Vision and Pattern Recognition, Radiology, Nuclear Medicine and Imaging

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Radiology, Nuclear Medicine and Imaging

Citations in the corpus

You Only Look at One Sequence: Rethinking Transformer in Vision through Object Detection (2021, methodology): “From the model scaling perspective [20, 56, 61], YOLOS-Ti can serve as a promising model scaling start point.”
Simple Training Strategies and Model Scaling for Object Detection (2021, neutral): “State-of-the-art object detection performance on the COCO benchmark [25] has been pushed from 30% AP to 58% since the development of popular object detectors [27, 29, 30, 12, 11, 31, 13, 3, 6, 24, 40, 23, 38, 26, 10, 8, 7, 36, 4, 2, 39, 9, 37, 28].”
YOLOX: Exceeding YOLO Series in 2021 (2021, result): “We notice that there are some high performance YOLO series with larger model sizes like Scale-YOLOv4 [30] and YOLOv5-P6 [7].”
TPH-YOLOv5: Improved YOLOv5 Based on Transformer Prediction Head for Object Detection on Drone-captured Scenarios (2021, methodology): “3) anchor-based detectors: Scaled-YOLOv4 [51], YOLOv5 [21].”
No More Strided Convolutions or Pooling: A New CNN Building Block for Low-Resolution Images and Small Objects (2022, neutral): “Faster R-CNN [27], SSD [24], RetinaNet [21], EﬃcientDet [34], YOLO [26,4,14,36] FCOS [35], CenterNet [7], DETR [5], YOLOX [8] Two-stage R-CNN [10], Fast R-CNN [9] RepPoints, CenterNet2 Fig.”
DETRs Beat YOLOs on Real-time Object Detection (2023, background): “YOLO detectors can be classified into two categories: anchor-based [1, 10, 14, 24, 27, 28, 34, 35] and anchor-free [9, 11, 15, 37], which achieve a reasonable trade-off between speed and accuracy and are widely used in various practical scenarios.”
Hyper-YOLO: When Visual Object Detection Meets Hypergraph Computation (2024, neutral): “The YOLO series methods [1], [2], [3], [4], [5], [7], [8], [22], [47], [48], [49], [50], [51] are typically composed of two main components: backbone and neck.”
PP-YOLOE: An evolved version of YOLO (2022, background): “2% YOLOv4-CSP [26] Modified CSPDarknet-53 512 97 - 46.”
Simple Multi-dataset Detection (2021, background): “This fragmentation brought rapid progress in object detection [5, 10, 20, 31, 39, 45] and instance segmentation [14], but comes with a drawback.”
A Survey of Modern Deep Learning based Object Detection Models (2021, methodology): “It is easy to implement and general enough to be applicable on other architectures like ResNet [21], ResNeXt [24], DenseNet [23], Scaled-YOLOv4 [26] etc. Applying CSPNet on these networks reduced computations from 10% to 20%, while the accuracy remained constant or improved.”
A Comprehensive Review of YOLO Architectures in Computer Vision: From YOLOv1 to YOLOv8 and YOLO-NAS (2023, background): “10 Scaled-YOLOv4 One year after YOLOv4, the same authors presented Scaled-YOLOv4 [87] in CVPR 2021.”
CBNet: A Composite Backbone Network Architecture for Object Detection (2021, neutral): “8% AP, outperforming previous anchor-free methods [6], [25], [26], [27], [30], [34], [38], including their multi-scale testing results; for anchor-based detectors [22], [23], [31], [32], [33], [39], [40], [41], our Cascade R-CNN DualRes2Net101-DCN achieves a comparable result of 55.”
YOLOv10: Real-Time End-to-End Object Detection (2024, methodology): “Besides, model scaling strategies [62, 61] and re-parameterization [11, 29] techniques are also investigated.”
You Only Learn One Representation: Unified Network for Multiple Tasks (2021, methodology): “Due to the stem block contains three down-sampling modules, YOLOv4-CSP-SSSS belongs to topology IV.”
Designing Network Design Strategies Through Gradient Path Analysis (2022, background): “One of the best examples is scaled-YOLOv4 [31], we see that its P7 model uses expensive parameters and operations, but only a small amount of accuracy gain, and the same phenomenon occurs in many popular networks.”
DAMO-YOLO : A Report on Real-Time Object Detection Design (2022, background): “Recently, researchers have developed object detection methods in huge progress [1,9,11,23,24,27,33].”