Main Content


Train YOLO v4 object detector

Since R2022a



detector = trainYOLOv4ObjectDetector(trainingData,detector,options) returns an object detector trained using you only look once version 4 (YOLO v4) network specified by the input detector. The input detector can be an untrained or pretrained YOLO v4 object detector. The options input specifies training parameters for the detection network.

You can also use this syntax for fine-tuning a pretrained YOLO v4 object detector.

detector = trainYOLOv4ObjectDetector(trainingData,checkpoint,options) resumes training from the saved detector checkpoint.

You can use this syntax to:

  • Add more training data and continue the training.

  • Improve training accuracy by increasing the maximum number of iterations.

[detector,info] = trainYOLOv4ObjectDetector(___) also returns information on the training progress, such as the training accuracy and learning rate for each iteration.

___ = trainYOLOv4ObjectDetector(___,Name=Value) specifies options using one or more name-value arguments in addition to any combination of arguments from previous syntaxes. For example, trainYOLOv4ObjectDetector(trainingData,ExperimentManager="none") sets the metrics to track with Experiment Manager to "none".


To run this function, you will require the Deep Learning Toolbox™.


collapse all

This example shows how to fine-tune a pretrained YOLO v4 object detector for detecting vehicles in an image. This example uses a tiny YOLO v4 network trained on COCO dataset.

Load a pretrained YOLO v4 object detector and inspect its properties.

detector = yolov4ObjectDetector("tiny-yolov4-coco")
detector = 
  yolov4ObjectDetector with properties:

        Network: [1×1 dlnetwork]
    AnchorBoxes: {2×1 cell}
     ClassNames: {80×1 cell}
      InputSize: [416 416 3]
      ModelName: 'tiny-yolov4-coco'

The number of anchor boxes must be same as that of the number of output layers in the YOLO v4 network. The tiny YOLO v4 network contains two output layers.

ans = 
  dlnetwork with properties:

         Layers: [74×1 nnet.cnn.layer.Layer]
    Connections: [80×2 table]
     Learnables: [80×3 table]
          State: [38×3 table]
     InputNames: {'input_1'}
    OutputNames: {'conv_31'  'conv_38'}
    Initialized: 1

Prepare Training Data

Load a .mat file containing information about the vehicle dataset to use for training. The information stored in the .mat file is a table. The first column contains the training images and the remaining columns contain the labeled bounding boxes.

data = load("vehicleTrainingData.mat");
trainingData = data.vehicleTrainingData;

Specify the directory in which training samples are stored. Add full path to the file names in training data.

dataDir = fullfile(toolboxdir("vision"),"visiondata");
trainingData.imageFilename = fullfile(dataDir,trainingData.imageFilename);

Create an imageDatastore using the files from the table.

imds = imageDatastore(trainingData.imageFilename);

Create a boxLabelDatastore using the label columns from the table.

blds = boxLabelDatastore(trainingData(:,2:end));

Combine the datastores.

ds = combine(imds,blds);

Specify the input size to use for resizing the training images. The size of the training images must be a multiple of 32 for when you use the tiny-yolov4-coco and csp-darknet53-coco pretrained YOLO v4 deep learning networks. You must also resize the bounding boxes based on the specified input size.

inputSize = [224 224 3];

Resize and rescale the training images and the bounding boxes by using the preprocessData helper function. Also, convert the preprocessed data to a datastore object by using the transform function.

trainingDataForEstimation = transform(ds,@(data)preprocessData(data,inputSize));

Estimate Anchor Boxes

Estimate the anchor boxes from the training data. You must assign the same number of anchor boxes to each output layer in the YOLO v4 network.

numAnchors = 6;
[anchors, meanIoU] = estimateAnchorBoxes(trainingDataForEstimation,numAnchors);
area = anchors(:,1).*anchors(:,2);
[~,idx] = sort(area,"descend");
anchors = anchors(idx,:);
anchorBoxes = {anchors(1:3,:);anchors(4:6,:)};

Configure and Train YOLO v4 Network

Specify the class names and configure the pretrained YOLO v4 deep learning network to be retrained for the new dataset by using yolov4ObjectDetector function.

classes = {"vehicle"};
detector = yolov4ObjectDetector("tiny-yolov4-coco",classes,anchorBoxes,InputSize=inputSize);

Specify the training options and retrain the pretrained YOLO v4 network on the new dataset by using the trainYOLOv4ObjectDetector function.

options = trainingOptions("sgdm", ...
    InitialLearnRate=0.001, ...
    MaxEpochs=40, ...
trainedDetector = trainYOLOv4ObjectDetector(ds,detector,options);
Training a YOLO v4 Object Detector for the following object classes:

* vehicle

    Epoch    Iteration    TimeElapsed    LearnRate    TrainingLoss
    _____    _________    ___________    _________    ____________
      2         30         00:01:07        0.001         7.215    
      4         60         00:01:44        0.001         1.7371   
      5         90         00:02:21        0.001        0.97954   
      7         120        00:02:57        0.001        0.59412   
      8         150        00:03:34        0.001        0.65631   
     10         180        00:04:10        0.001         1.0774   
     12         210        00:04:46        0.001         0.4807   
     13         240        00:05:22        0.001        0.40389   
     15         270        00:05:59        0.001        0.57931   
     16         300        00:06:35        0.001        0.90734   
     18         330        00:07:11        0.001        0.24902   
     19         360        00:07:48        0.001        0.32441   
     21         390        00:08:24        0.001        0.23054   
     23         420        00:09:00        0.001        0.70897   
     24         450        00:09:36        0.001        0.31744   
     26         480        00:10:12        0.001        0.36323   
     27         510        00:10:49        0.001        0.13696   
     29         540        00:11:25        0.001        0.14913   
     30         570        00:12:01        0.001        0.37757   
     32         600        00:12:37        0.001        0.36985   
     34         630        00:13:14        0.001        0.14034   
     35         660        00:13:50        0.001        0.14731   
     37         690        00:14:26        0.001        0.15907   
     38         720        00:15:03        0.001        0.11737   
     40         750        00:15:40        0.001         0.1855   

Detector training complete.

Detect Vehicles in Test Image

Read a test image.

I = imread("highway.png");

Use the fine-tuned YOLO v4 object detector to detect vehicles in a test image and display the detection results.

[bboxes, scores, labels] = detect(trainedDetector,I,Threshold=0.05);
detectedImg = insertObjectAnnotation(I,"Rectangle",bboxes,labels);

function data = preprocessData(data,targetSize)
for num = 1:size(data,1)
    I = data{num,1};
    imgSize = size(I);
    bboxes = data{num,2};
    I = im2single(imresize(I,targetSize(1:2)));
    scale = targetSize(1:2)./imgSize(1:2);
    bboxes = bboxresize(bboxes,scale);
    data(num,1:2) = {I,bboxes};

Input Arguments

collapse all

Labeled ground truth images, specified as a datastore. Data must be set up so that calling the datastore with the read and readall functions returns a cell array or table with three columns in the format {data,boxes,labels}.

The first column, data, must contain the image data, stored as a cell array. The second column, boxes, must contain the bounding boxes. The third column, labels, must be a cell array that contains M-by-1 categorical vectors containing object class names, where M is the number of bounding boxes. All the categorical data returned by the datastore must use the same categories.

The table describes the format of the bounding boxes column.

Bounding BoxDescription

Axis-aligned rectangle

Defined in spatial coordinates as an M-by-4 numeric matrix with rows of the form [x y w h], where:

  • M is the number of axis-aligned rectangles.

  • x and y specify the upper-left corner of the rectangle.

  • w specifies the width of the rectangle, which is its length along the x-axis.

  • h specifies the height of the rectangle, which is its length along the y-axis.

Rotated rectangle

Defined in spatial coordinates as an M-by-5 numeric matrix with rows of the form [xctr yctr w h yaw], where:

  • M is the number of rotated rectangles.

  • xctr and yctr specify the center of the rectangle.

  • w specifies the width of the rectangle, which is its length along the x-axis before rotation.

  • h specifies the height of the rectangle, which is its length along the y-axis before rotation.

  • yaw specifies the rotation angle in degrees. The rotation is clockwise-positive around the center of the bounding box.

Square rectangle rotated by -30 degrees.

For more information, see Datastores for Deep Learning (Deep Learning Toolbox).


A pretrained axis-aligned network can be converted to a rotated rectangle network by providing rotated rectangle training data. When you provide the rotated rectangle training data, the trainYOLOv4ObjectDetector function fine tunes the network heads allowing the rotated rectangle detections to occur.

Pretrained or untrained YOLO v4 object detector, specified as a yolov4ObjectDetector object.

Training options, specified as a TrainingOptionsSGDM, TrainingOptionsRMSProp, or TrainingOptionsADAM object returned by the trainingOptions (Deep Learning Toolbox) function. To specify the solver name and other options for network training, use the trainingOptions function.

Saved detector checkpoint, specified as a yolov4ObjectDetector object. To periodically save a detector checkpoint during training, specify CheckpointPath. To control how frequently check points are saved see the CheckPointFrequency and CheckPointFrequencyUnit training options.

To load a checkpoint for a previously trained detector, load the MAT-file from the checkpoint path. For example, if the CheckpointPath property of the object specified by options is 'checkpath', you can load a checkpoint MAT-file by using this code. 'checkpath' is the name of a folder in the current working directory to which the detector checkpoint has to be saved during training.

data = load('checkpath/net_checkpoint__19__2021_12_29__01_04_15.mat');
checkpoint =;

The name of the MAT-file includes the iteration number and timestamp of when the detector checkpoint was saved. The detector is saved in the net variable of the file. Pass this file back into the trainYOLOv4ObjectDetector function:

yoloDetector = trainYOLOv4ObjectDetector(trainingData,checkpoint,options);

Name-Value Arguments

Specify optional pairs of arguments as Name1=Value1,...,NameN=ValueN, where Name is the argument name and Value is the corresponding value. Name-value arguments must appear after other arguments, but the order of the pairs does not matter.

Before R2021a, use commas to separate each name and value, and enclose Name in quotes.

Example: ExperimentManager="none" sets the metrics to track with Experiment Manger to "none".

Since R2024a

Subnetworks to freeze during training, specified as one of these values:

  • "none" — Do not freeze subnetworks

  • "backbone" — Freeze the feature extraction subnetwork

  • "backboneAndNeck" — Freeze both the feature extraction and the path aggregation subnetworks

The weight of layers in frozen subnetworks does not change during training.


You cannot use the FreezeSubNetwork argument values "backbone" and "backboneAndNeck" with a custom YOLO v4 object detector created using the syntax yolov4ObjectDetector(net,classes,aboxes).

Detector training experiment monitoring, specified as an experiments.Monitor (Deep Learning Toolbox) object for use with the Experiment Manager (Deep Learning Toolbox) app. You can use this object to track the progress of training, update information fields in the training results table, record values of the metrics used by the training, and to produce training plots. For an example using this app, see Train Object Detectors in Experiment Manager.

Information monitored during training:

  • Training loss at each iteration.

  • Learning rate at each iteration.

Validation information when the training options input contains validation data:

  • Validation loss at each iteration.

Output Arguments

collapse all

Trained YOLO v4 object detector, returned as yolov4ObjectDetector object. You can train a YOLO v4 object detector to detect multiple object classes.

Training progress information, returned as a structure array with seven fields. Each field corresponds to a stage of training.

  • TrainingLoss — Training loss at each iteration. The trainYOLOv4ObjectDetector function uses mean square error for computing bounding box regression loss and cross-entropy for computing classification loss.

  • BaseLearnRate — Learning rate at each iteration.

  • OutputNetworkIteration — Iteration number of returned network.

  • ValidationLoss — Validation loss at each iteration.

  • FinalValidationLoss — Final validation loss at end of the training.

Each field is a numeric vector with one element per training iteration. Values that have not been calculated at a specific iteration are assigned as NaN. The struct contains ValidationLoss and FinalValidationLoss fields only when options specifies validation data.


  • To generate the ground truth, use the Image Labeler or Video Labeler app. To create a table of training data from the generated ground truth, use the objectDetectorTrainingData function.

  • To improve prediction accuracy,

    • Increase the number of images you can use to train the network. You can expand the training dataset through data augmentation. For information on how to apply data augmentation for preprocessing, see Preprocess Images for Deep Learning (Deep Learning Toolbox).

    • Choose anchor boxes appropriate to the dataset for training the network. You can use the estimateAnchorBoxes function to compute anchor boxes directly from the training data.

  • A pretrained axis-aligned network can be converted to a rotated rectangle network by providing rotated rectangle training data. When you provide the rotated rectangle training data, the trainYOLOv4ObjectDetector function fine tunes the network heads allowing the rotated rectangle detections to occur.

  • When you train a rotated rectangle bounding box detector, use a learning rate approximately one order of magnitude below that of its axis-aligned counterpart training rate.

  • When you perform transfer learning using a YOLO v4 object detector, consider freezing the subnetworks using the name-value argument FreezeSubNetwork to increase training speed and reduce GPU memory consumption.

Version History

Introduced in R2022a

expand all