• YOLOv4: https://arxiv.org/abs/2004.10934

• Scaled-YOLOv4: https://openaccess.thecvf.com/content/CVPR2021/html/Wang_Scaled-YOLOv4_Scaling_Cross_Stage_Partial_Network_CVPR_2021_paper.html

• YOLOR: https://arxiv.org/abs/2105.04206

YOLOv4

YOLOv4 consists of: https://arxiv.org/abs/2004.10934

• Backbone: CSPDarknet53
• Neck: SPP, PAN
• Head: YOLOv3

Download:

• cfg-file: https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v3_optimal/yolov4.cfg

• weights-file: https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v3_optimal/yolov4.weights

• pre-trained weights-file for training: https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v3_optimal/yolov4.conv.137

Accuracy (AP) / Speed (FPS):

YOLOv4 (416x416 batch=1 FP16) - 32 FPS on Jetson Xavier AGX by using tkDNN+TensorRT: https://github.com/ceccocats/tkDNN

Features:

• fusion blocks: FPN, PAN, ASFF, BiFPN
• network modules: ResNet, CPS, SPP, RFB
• network architecture search: CSPResNext50, CSPDarknet53, SpineNet49, EfficientNetB0, MixNet-M
• activations: SWISH, MISH
• other features: weighted-[shortcut], Sigmoid scaling (Scale-sensitivity), Label smoothing, Optimal hyper parameters, Dynamic mini batch size for random shapes, Squeeze-and-excitation, Grouped convolution, MixConv (grouped [route]), Elastic-module
• data augmentation: MixUp, CutMix, Mosaic
• losses: MSE, GIoU, CIoU, DIoU
• detection layers: [yolo] (fixed iou_thresh), [Gaussian_yolo]
• detection on video (sequence of frames) - layers: [crnn] (convolutional-RNN), [conv_lstm] (Convolutional LSTM)

Darknet for Windows & Linux:

• object detection: Yolo v3, Yolo v2
• classification: ResNet, Darknet, DenseNet
• sequence prediction RNN-layers: rnn, crnn, gru, lstm
• Tensor Cores are used

Tested for Training and Prediction.

• There are attached compiled binary files of Darknet for Windows x64 (559 MB): https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v3/darknet.zip

  Are used: MSVS 2015, CUDA 10.0, cuDNN 7.4.2, OpenCV 3.4.0, GPU CC 3.0 & 7.5
  May be required vc_redist.x64.exe file 13.9 MB (Visual C++ Redistributable for MSVS2015 ): https://www.microsoft.com/en-us/download/confirmation.aspx?id=48145&6B49FDFB-8E5B-4B07-BC31-15695C5A2143=1

Can be used for training and detection Yolo v3 and v2 on Windows and Linux.
Added several performance improvements.
It supports mixed-precision training/detection using Tensor Cores on GPU Volta - set CUDNN_HALF=1 in the Makefile