YOLOv4 is an upgraded version from YOLOv3. YOLO is the abbreviation of You Look Only Once neural network technology. It is the very famous real-time Object Recognition technology that is capable of recognising multiple objects in a single frame. Market YOLOv4 is designed with the efficiency of identifying objects at more accuracy and speed than every other system exists in the market like Google TensorFlow, RetinaNet/MaskRCNN, EfficientDet, and FaceBook Pytorch/Detectron, Yolov3-ASFF on Microsoft COCO dataset. It is capable of recognising more than 9000 classes along with undetectable classes. It could also specify a boundary box around the object. YOLOv4 is highly flexible that can be easily deployed in any production environment.

YOLOv4 is found to be much efficient that it can achieve 43.5% Average Precision(AP)/ 65.7% AP50 accuracy in accordance with the Microsoft COCO test and is at the fastest speed of 62 Frames Per Second (FPS) TitanV or 34 FPS RTX 2070.

Basically, Yolo is built with a single convolution neural network (CNN)algorithm, which leads the market in making a faster and efficient computer vision algorithms.

Now not only the mighty players can train a neural network algorithm, even the individual users can achieve training and testing CNN algorithms with the help of YOLOv4 by simply using the nVidia adapter for gaming graphics with 8–16 GB VRAM. This cannot be done with the previous contemporary detectors.

Now let's look into the architecture and improvements in YoloV4.

YoloV4 Architecture:

• CSPDarknet53 as the backbone for the structure to increase the capability of CNN

• Spatial Pyramid Pooling is an additional module to increase the reception field to segregate the context features

• PANet path-aggregation is used to collaborate the various detector classes

• YOLOv3 Head, the source architecture where these additional characteristics are appended.

Functional Improvements in YOLOv4:

• It is considered to be twice as fast as in all parameters when compared to that of the existing competing object recognition models, such as EfficientDet, etc.

• Average Precision(AP) is a crucial parameter to measure the accuracy in the real-time object recognition is found to increase by 10% than the available models.

• Frames Per Second(FPS) is to measure the speed and if found to be increased to 12% in YOLOv4 with that of the YOLOv3

References:

PDF: YOLOv4: Optimal Speed and Accuracy of Object Detection

Github: Yolo-v4 Darknet

Github: Yolo-v4 and Yolo-v3/v2 for Windows and Linux