yolov3ObjectDetector

Detect objects using YOLO v3 object detector

Since R2021a

Description

The yolov3ObjectDetector object creates a you only look once version 3 (YOLO v3) object detector for detecting objects in an image. Using this object, you can:

• Create a pretrained YOLO v3 object detector by using YOLO v3 deep learning networks trained on COCO dataset.

• Create a custom YOLO v3 object detector by using any pretrained or untrained YOLO v3 deep learning network.

Creation

Syntax

detector = yolov3ObjectDetector(name)

detector = yolov3ObjectDetector(name,classes,aboxes)

detector = yolov3ObjectDetector(net,classes,aboxes)

detector = yolov3ObjectDetector(baseNet,classes,aboxes,DetectionNetworkSource=layer)

detector = yolov3ObjectDetector(___,Name=Value)

Description

Pretrained YOLO v3 Object Detector

detector = yolov3ObjectDetector(name) creates a pretrained YOLO v3 object detector by using YOLO v3 deep learning networks trained on a COCO dataset.

Note

To use the pretrained YOLO v3 deep learning networks trained on COCO dataset, you must install the Computer Vision Toolbox™ Model for YOLO v3 Object Detection from Add-On Explorer. For more information about installing add-ons, see Get and Manage Add-Ons. To run this function, you will require the Deep Learning Toolbox™.

Custom YOLO v3 Object Detector

detector = yolov3ObjectDetector(name,classes,aboxes) creates a pretrained YOLO v3 object detector and configures it to perform transfer learning using a specified set of object classes and anchor boxes. For optimal results, you must train the detector on new training images before performing detection.

detector = yolov3ObjectDetector(net,classes,aboxes) creates an object detector by using the deep learning network net.

If net is a pretrained YOLO v3 deep learning network, the function creates a pretrained YOLO v3 object detector. The classes and aboxes are values used for training the network.

If net is an untrained YOLO v3 deep learning network, the function creates a YOLO v3 object detector to use for training and inference. classes and aboxes specify the object classes and the anchor boxes, respectively, for training the YOLO v3 network.

You must train the detector on a training dataset before performing object detection. For information about how to train a YOLO v3 object detector, see Preprocess Training Data and Train Model sections in the Object Detection Using YOLO v3 Deep Learning example.

detector = yolov3ObjectDetector(baseNet,classes,aboxes,DetectionNetworkSource=layer) creates a YOLO v3 object detector by adding detection heads to a base network, baseNet.

The function adds detection heads to the specified feature extraction layers layer in the base network. To specify the names of the feature extraction layers, use the name-value argument DetectionNetworkSource=layer.

If baseNet is a pretrained deep learning network, the function creates a YOLO v3 object detector and configures it to perform transfer learning with the specified object classes and anchor boxes.

If baseNet is an untrained deep learning network, the function creates a YOLO v3 object detector and configures it for object detection. classes and aboxes specify the object classes and the anchor boxes, respectively, for training the YOLO v3 network.

You must train the detector on a training dataset before performing object detection.

detector = yolov3ObjectDetector(___,Name=Value) specifies one or more options using name-value arguments in addition to any combination of input arguments from previous syntaxes. Use this syntax to

• Modify the detection network sources in a yolo v4 object detection network and train the network with different numbers of object classes, anchor boxes, or both. Specify the new detection network sources using the name-value argument DetectionNetworkSource=layer.

• Set the InputSize and ModelName properties of the object detector. For example, InputSize=[224 224 3] sets the size of the images used for training to [224 224 3].

Input Arguments

name — Name of pretrained YOLO v3 deep learning network

'darknet53-coco' | 'tiny-yolov3-coco'

Name of the pretrained YOLO v3 deep learning network, specified as one of these:

• 'darknet53-coco' — A pretrained YOLO v3 deep learning network created using DarkNet-53 as the base network and trained on COCO dataset.

• 'tiny-yolov3-coco' — A pretrained YOLO v3 deep learning network created using a small base network and trained on COCO dataset.

Data Types: char | string

classes — Names of object classes

string vector | cell array of character vectors | categorical vector

Names of object classes for training the detector, specified as a string vector, cell array of character vectors, or categorical vector. This argument sets the ClassNames property of the yolov3ObjectDetector object.

Data Types: char | string | categorical

aboxes — Anchor boxes

N-by-1 cell array

Anchor boxes for training the detector, specified as an N-by-1 cell array. N is the number of output layers in the YOLO v3 deep learning network. Each cell contains an M-by-2 matrix, where M is the number of anchor boxes in that layer. Each cell can contain a different number of anchor boxes. Each row in the M-by-2 matrix denotes the size of an anchor box in the form [height width].

The first element in the cell array specifies the anchor boxes to associate with the first output layer, the second element in the cell array specifies the anchor boxes to associate with the second output layer, and so on. For accurate detection results, specify large anchor boxes for the first output layer and small anchor boxes for the last output layer. That is, the anchor box sizes must decrease for each output layer in the order in which the layers appear in the YOLO v3 deep learning network.

This argument sets the AnchorBoxes property of the yolov3ObjectDetector object.

Data Types: cell

net — YOLO v3 deep learning network

dlnetwork object

YOLO v3 deep learning network, specified as a dlnetwork (Deep Learning Toolbox) object. The input network can be either an untrained or a pretrained deep learning network.

baseNet — Base network

dlnetwork object

Base network for creating the YOLO v3 deep learning network, specified as a dlnetwork (Deep Learning Toolbox) object. The network can be either an untrained or a pretrained deep learning network.

layer — Names of feature extraction layers

cell array of character vectors | string array

Names of the feature extraction layers in the base network, specified as a cell array of character vectors, or a string array. The function creates a YOLO v3 network by adding detection head layers to the output of the feature extraction layers in the base network.

In the pre-trained YOLO v3 base network, you must connect the detection head layers to feature extraction layers where the input image feature map changes spatial dimension by downsampling factors of 4, 8, 16, and 32.

• Use the analyzeNetwork (Deep Learning Toolbox) function to display the pre-trained YOLO v3 network architecture to obtain the information about the spatial dimensions of the feature maps.

• Use this table to know about the default feature extraction layers and choose alternate feature extraction layers supported by the pre-trained YOLO v3 networks. Choose feature extraction layers using empirical evaluation based on the detection accuracy, training speed, and object size to detect.

  • For multi-scale feature extraction and to achieve higher detection accuracy, choose feature extraction layers with both higher and lower downsampling factors. Layers with lower downsampling factor capture small objects and layers with higher downsampling factor ensure a thorough feature refinement.

  • To achieve higher training and inference speeds, choose feature extraction layers with lower downsampling factor.

  • The number of detection network sources specified must be equal to the number of anchor boxes specified at the input.

  Backbone	Number of Detection Heads	Default Feature Extraction Layers		Examples of Alternate Feature Extraction Layers		Remarks
  		Layer	Downsampling Factor	Layer	Downsampling Factor
  tiny-yolov3-coco	2	leaky_relu_7 leaky_relu_5	32 16	leaky_relu_6 leaky_relu_3	16 8	Select any two feature extraction layers with different downsampling factors preceding the leaky_relu_7 layer to use as detection network sources. For example, specify the new detection network sources as layer = {'leaky_relu_6','leaky_relu_3'}
  darknet53-coco	3	add_23 add_19 add_11	32 16 8	add_22 add_21 add_20	32	Select any three feature extraction layers with different downsampling factors preceding the add_23 layer to use as detection network sources. For example, specify the new detection network sources as layer = {'add_18','add_10','add_3'}
  				add_18 add_17 add_16 add_15 add_14 add_13 add_12	16
  				add_10 add_9 add_8 add_7 add_6 add_5 add_4	8
  				add_3 add_2	4

Data Types: char | string | cell

Properties

Network — YOLO v3 deep learning network

dlnetwork object

This property is read-only.

YOLO v3 deep learning network to use for object detection, stored as a dlnetwork (Deep Learning Toolbox) object.

AnchorBoxes — Set of anchor boxes

N-by-1 cell array

This property is read-only.

Set of anchor boxes, stored as a N-by-1 cell array. N is the number of output layers in the YOLO v3 deep learning network for which the anchor boxes are defined. Each element in the cell is a M-by-2 matrix. M denotes the number of anchor boxes. Each cell can contain a different number of anchor boxes. Each row in the M-by-2 matrix denotes the size of the anchor box in the form of [height width]. The first element in the cell array specifies the anchor boxes for the first output layer, the second element in the cell array specifies the anchor boxes for the second output layer, and so on.

You can set this property by using the input argument aboxes.

ClassNames — Names of object classes

categorical vector

This property is read-only.

Names of object classes to detect, stored as a categorical vector. You can set this property by using the input argument classes.

InputSize — Image size used for training

vector

This property is read-only.

Image size used for training, stored as a vector of form [height width] or [height width channels]. To set this property, specify it at object creation.

For example, detector = yolov3ObjectDetector(net,classes,aboxes,InputSize=[224 224 3]).

ModelName — Name for object detector

' ' (default) | character vector

Name for the object detector, stored as a character vector. To set this property, specify it at object creation.

For example, yolov3ObjectDetector(net,classes,aboxes,'ModelName','customDetector') sets the name for the object detector to 'customDetector'.

PredictedBoxType — Bounding box format for object detector

"axis-aligned" (default) | "rotated"

This property is read-only.

Bounding box format for an object detector, stored as "axis-aligned" or "rotated". When the PredictedBoxType is "axis-aligned", the object detector will train and perform inference on only axis-aligned bounding boxes. If it is set to "rotated", the object detector will train and perform inference on only rotated bounding boxes. Set this property when you create the object.

Object Functions

detect	Detect objects using YOLO v3 object detector
preprocess	Preprocess training and test images
forward	Compute YOLO v3 deep learning network output for training
predict	Compute YOLO v3 deep learning network outputs for inference

Examples

Create Pretrained YOLO v3 Object Detector

Specify the name of a pretrained YOLO v3 deep learning network.

name = 'tiny-yolov3-coco';

Create YOLO v3 object detector by using the pretrained YOLO v3 network.

detector = yolov3ObjectDetector(name);

Display and inspect the properties of the YOLO v3 object detector.

disp(detector)

  yolov3ObjectDetector with properties:

             Network: [1x1 dlnetwork]
         AnchorBoxes: {2x1 cell}
          ClassNames: [80x1 categorical]
           InputSize: [416 416 3]
    PredictedBoxType: 'axis-aligned'
           ModelName: 'tiny-yolov3-coco'

Use analyzeNetwork to display the YOLO v3 network architecture and get information about the network layers. The network has two detection heads attached to the feature extraction network.

analyzeNetwork(detector.Network)

Detect objects in an unknown image by using the pretrained YOLO v3 object detector.

img = imread('sherlock.jpg');
img = preprocess(detector,img);
img = im2single(img);
[bboxes,scores,labels] = detect(detector,img,'DetectionPreprocessing','none');

Display the detection results.

detectedImg = insertObjectAnnotation(img,'Rectangle',bboxes,labels);
figure
imshow(detectedImg)

Create Custom YOLO v3 Object Detector

This example shows how to create a custom YOLO v3 object detector by using a pretrained SqueezeNet as the base network.

Load a pretrained SqueezeNet network. The SqueezeNet network is a convolutional neural network that you can use as the base network for creating a YOLO v3 object detector.

net = imagePretrainedNetwork("squeezenet");

Inspect the architecture of the base network by using analyzeNetwork (Deep Learning Toolbox) function.

analyzeNetwork(net)

Specify the anchor boxes and the classes to use to train the YOLO v3 network.

aboxes = {[150,127;97,90;68,67];[38,42;41,29;31,23]};
classes = ["Car","Truck"];

Select two feature extraction layers in the base network to serve as the source for detection subnetwork.

layer = ["fire9-concat","fire8-concat"];

Create a custom YOLO v3 object detector by adding detection heads to the feature extraction layers of the base network. Specify the model name, classes, and the anchor boxes.

detector = yolov3ObjectDetector(net,classes,aboxes, ...
    ModelName="Custom YOLO v3",DetectionNetworkSource=layer);

Inspect the architecture of the YOLO v3 deep learning network by using analyzeNetwork (Deep Learning Toolbox) function.

analyzeNetwork(detector.Network)

Inspect the properties of the YOLO v3 object detector. You can now train the YOLO v3 object detector on a custom training dataset and perform object detection.

disp(detector)

  yolov3ObjectDetector with properties:

             Network: [1x1 dlnetwork]
         AnchorBoxes: {2x1 cell}
          ClassNames: [2x1 categorical]
           InputSize: [227 227 3]
    PredictedBoxType: 'axis-aligned'
           ModelName: 'Custom YOLO v3'

For information about how to train a YOLO v3 object detector, see the Object Detection Using YOLO v3 Deep Learning example.

Extended Capabilities

C/C++ Code Generation

Generate C and C++ code using MATLAB® Coder™.

Usage notes and limitations:

• Only the detect method of the yolov3ObjectDetector is supported for code generation.

• The roi argument to the detect method must be a code generation constant (coder.const()) and a 1x4 vector.

• Only the Threshold, SelectStrongest, MinSize, and MaxSize name-value pairs for detect are supported.

• To create a yolov3ObjectDetector object for code generation, see Load Pretrained Networks for Code Generation (MATLAB Coder).

GPU Code Generation

Generate CUDA® code for NVIDIA® GPUs using GPU Coder™.

Usage notes and limitations:

• Only the detect method of the yolov3ObjectDetector is supported for code generation.

• The roi argument to the detect method must be a codegen constant (coder.const()) and a 1x4 vector.

• Only the Threshold, SelectStrongest, MinSize, and MaxSize name-value pairs are supported.

• The height, width, and channel of the input image must be fixed size.

• To create a yolov3ObjectDetector object for code generation, see Load Pretrained Networks for Code Generation (MATLAB Coder).

GPU Arrays

Accelerate code by running on a graphics processing unit (GPU) using Parallel Computing Toolbox™.

Usage notes and limitations:

• Only the detect method of the yolov3ObjectDetector is supported for GPU Arrays based processing.

• GPU Arrays is not supported for rotated rectangle bounding box inputs.

Version History

Introduced in R2021a

R2024b: Code generation support for rotated rectangle bounding boxes

Starting in R2024b, YOLO v3 object detector with PredictedBoxType property set as "rotated" supports both CPU and GPU code generation.

R2024b: Modify detection network source

Starting in R2024b, you can modify the detection network sources of a YOLO v3 object detection network when you perform transfer learning. You can specify the new detection network sources using the name-value argument DetectionNetworkSource=layer.

• Specify new detection network sources for the tiny-yolov3-coco and darknet53-coco pretrained YOLO v3 object detectors by using the syntax

  detector = yolov3ObjectDetector(name,classes,aboxes,DetectionNetworkSource=layer);

• Specify new detection network sources for a YOLO v3 object detection network by using the syntax

  detector = yolov3ObjectDetector(net,classes,aboxes,DetectionNetworkSource=layer);

For an example, the below code modifies the detection network sources of the pretrained tiny-yolov3-coco network to leaky_relu_2 and leaky_relu_5.

name = "tiny-yolov3-coco";
layer = ["leaky_relu_2","leaky_relu_5"];
detector = yolov3ObjectDetector(name,classes,aboxes,DetectionNetworkSource=layer);

R2024a: DAGNetwork objects are not recommended

Starting in R2024a, DAGNetwork (Deep Learning Toolbox) objects are not recommended. If you specify a base network baseNet, use a dlnetwork (Deep Learning Toolbox) object instead.

There are no plans to remove support for DAGNetwork objects.

R2024a: Rotated rectangle bounding box support

The deep learning YOLO v3 object detector now supports rotated rectangle bounding boxes.