YOLOv8 Architecture Detailed Explanation - A Complete Breakdown

hello and welcome to this video in this video I will explain about YOLO architecture before explaining the whole YOLO f8 architecture I will explain details of YOLO f8 blocks so that you can understand more easily the most commonly use block is the convolutional Block in yo8 a convolutional block consists of a 2d convolutional layer a 2d beds normalization and a seu activation function they are all fused together into a single convolutional block next one is the c2f block this block contains a convolutional block which then the resulting feature Maps will be split one goes to the bottleneck block whereas the other goes directly into the concet Block in c2f Block we can have many bottleneck blocks at the end there is another convolutional block butter neck itself is a sequence of convolutional blocks with a shortcut if you are familiar with ret block Bott neck block is pretty similar to rnet block the difference is that there is a bottom neck without a shortcut the SPF block is next SPF stands for spatial pyramid pooling fast it is a modification of SP of spatial pyramid pulling with a higher speed inside s ppf there are convolutional block at the beginning and followed by 32d Max pulling layer the interesting part is that every resulting F map is concatenated right before the end of SPF SPF is ending with a convolution of block the last block is the detect block this is where the detection happens different from the previous yolow version YOLO is an anchor-free model the predictions happen in the grid cell the detect block contains two tracks the first track is for bonding box prediction whereas the other is for class prediction both tracks has the same block sequence which is two convolutional blocks and a single 2D convolutional layer allowing that I will explain some of the fundamental components of a convolutional neural network first there is the kernel the kernel is a two-dimensional array kernel are usually called feature detectors the value in the kernel is weights which can be updated during the training process kernel will move across the image and perform a DOT operation between the input and the value of the kernel to produce an output the output is also known as a feat map next is the stride The Stride is defined as the displacement distance during the convolution process the smaller the resulting output the larger the strip convolution with stride one is demonstrated in this example then there is the pading padding is adding value to the uttermost element of the image in pyth words there are several types of pading first there is SOS padding this is the default padding type in the seros padding the pended pick will have a value of zero and then there is replication padding in the replication padding the padded pixels will have the same value as the closest real pixel the padic corners will have the same value as the real Corners next the YOLO fiet architecture in general the YOLO architecture is divided into three parts there are the backbone NE and head backbone is the Deep learning architecture that basically acts as a feature extractor the neck combines the features acquired from the various layers of the backbone model the head predicts the classes and buring box regions which is the final output produced by the object detection model however the neck is not explicitly mentioned in Yolo f8 the term neck is only written in the official YOLO f8 documentation on the YOLO f8 architecture file yolo. yml there are only two parts the backbone and the head next I will explain the whole YOLO fet architecture this architecture drawing is based on YOLO fet architecture file YOLO fate. yl which is located in models f8 folder it is also heavily inspired by the drawing from range King a GitHub user who posted an issue in Yolo GitHub repository we made some modifications to the drawing to make it more readable and align with the yo8 source code itself the explanation of architecture begins with an explanation of the three parameters that Define the YOLO Vari these parameters are depth multiple with multiple and Max channels the depth multiple parameters determine how many bottom neck blocks are in c2f Block the width multiple and Max channels parameters determine the output channel the YOLO at input is an image with three channels next the backbone the name of the backbone in Yolo is not stated directly on the backbone each backbone is made up of numerous convolution layers that extract distinct features at various resolution levels before continuing on the explanation of the layers on the backbone I will explain about the numbering on the YOLO f8 architecture each numbering is based on the architecture file which is yolo. yl numbering starts from the backbone section and starting from zero for example this convolution block is the first block in the architecture so we assign it to number zero and we draw the block is s below this numbering continues until the last c2f block this backbone begins with two convolutional blocks with kernel size three stride size two and padding one the spal resolution of the output is reduced when stride two is used for example if the input resolution in the first convolution of block is 640x 640 the output resolution after processing will be 320 by 320 to obtain the output Channel use the following formula this formula is obtained for the code in the tasks. pile first we find the minimum value between the base output Channel and Max channels the minimum value then multiplied by the width multiple parameters for example we will calculate the first convolution of blocks output Channel using the YOLO VL variant with a width multiple of one and a Max channels of 512 the base output channel in the first convolutional block is 64 so here is the calculation first we find the minimum value between 64 and 512 then multiply by 1 the result is 64 64 is the output channel in the first convolutional block if you use the YOLO you can analyze the second convolutional block with the same way as the first one next is the c2f block this block contains two parameters shortcut and N the shortcut parameters in this block is true indicating that the shortcut will be used on the botter neck block whereas n determines how many botter neck blocks are used the N value is calculated by multiplying the depth multiple value by three next there is another convolutional block with the kernel size of three strip two and padding one the c2f block comes next with the shortcut parameters through and in parameters equal to 6 multipli by the depth multiple this blocks output is also connected to the neck next there is another convolutional block with a kernel size of three stride two and padding one and then another c2f block with the shortcut parameters through and in parameters equal to 6 multip by the depth multiple this Block's output is also connected to the neck next there is another convolutional block with a kernel size of three stride two and padding one after that there is c2f block with the shortcut parameters through and the end parameters equal to three multipli by the depth multiple this block will be connected to SPF SPF spal pyramid pooling fast is used after the last convolution layer on the backbone the main function of the SPF is to generate a fixed fatal representation of objects of various sizes in an image without resizing the image or introducing special information loss following that an explanation of the neck first there is the upsample layer this layer is used to increase the feature map resolution of the SPF to match with the feature map resolution of this c2f block the up sample feature map will be combined with the features from this c2f block using concap when using concap the number of channels is summed up whereas the resolution is unchanged for example we will compute the concatenation of this c2f block feat map and this upsample feat M we use the YOLO variant the output of this c2f block is 40x 4 40x 512 and the up sample output is 40x 40x 512 the result of concatenation is 40x40 by 1,24 the following is c2f block on the neck c2f block does not imply a sort cut and the value of n equal to 3 mtip by the depth multiple the resolution of the c2f block fter map will be upsample to match the resolution of the Fe map of this c2f block using concap the upsample fure map will be combined with the features from this c2f block next there is another c2f block this block will reduce the channel size of the feat map the feat map of this block will be used as an input for the detect block this detect block is specialized for detecting small objects the output of this block is also used as an input to this convolutional block the convolutional block uses a kernel size of three stride two and padding one the resolution of the F map will will be reduced by half using this block furthermore conat will be used to combine the Fe map from this convolutional block with the Fe map from this c2f block next there is another c2f block this block will reduce the channel size of the feat map the feature map of this block will be used as input for the detect block this detect block is specialized for detecting medium siiz objects the output of this block is also used as input to this convolutional block the convolutional block uses a kernel size of three stride two and padding one next concat will be used to combine the F map from this convolutional block with the feat map from SPF block finally there is another c2f block this Block's feature map will be utilized as an input for the detect block this detect block is specialized for detecting large objects that's all the explanation about YOLO architecture eager to know how to greatly improve speed and accuracy by modifying YOLO architecture click the link in the description

Channel: Dr. Priyanto Hidayatullah

Views: 11,751

Keywords: yolo, yolov8, object detection, yolov8 accuracy, yolov8 speed, computer vision, deep learning

Id: HQXhDO7COj8

Length: 11min 6sec (666 seconds)

Published: Sat Oct 28 2023