create_dl_model_detection — Create a deep learning network for object detection.
create_dl_model_detection( : : Backbone, NumClasses, DLModelDetectionParam : DLModelHandle)
With the operator create_dl_model_detection a deep learning network
for object detection is created.
See the chapter Deep Learning / Object Detection for further
information on object detection based on deep learning.
The handle of this network is returned in DLModelHandle.
You can specify your model and its architecture over the parameters listed
below.
To successfully create a detection model, you need to specify its backbone
and the number of classes the model shall be able to distinguish.
The first information is handed over through the parameter
Backbone which is explained below in the section
“Possible Backbones”.
The second information is given through the parameter NumClasses.
Note, this parameter fixes the number of classes the network will
distinguish and therewith also the number of entries in
'class_ids'.
The values of all other applicable parameters can be specified using
the dictionary DLModelDetectionParam.
Such a parameter is e.g., the 'instance_type', determining which
kind of bounding boxes the model handles.
The full list of parameters that can be set is given below in the section
“Settable Parameters”.
In case a parameter is not specified, the default value is taken to create
the model.
Note, parameters influencing the network architecture will not be changeable
anymore once the network has been created.
All the other parameters can still be set or changed using the operator
set_dl_model_param.
An overview, how parameters can be set is given in
get_dl_model_param, where also a description of the specific
parameters is provided.
After creating the object detection model, the type will
automatically be set to 'detection'.
The parameter Backbone determines the backbone your
network will use. See the chapter Deep Learning / Object Detection
for more information to the backbone. In short, the backbone consists of a
pretrained classifier, from which only the layers necessary to generate
the feature maps are kept.
Hence, there are no fully connected layers anymore in the network.
This implies that you read in a classifier as feature extractor for
the subsequent detection network.
For this you can read in a classifier in the HALCON format or a model
or in the ONNX format, see read_dl_model for more information.
create_dl_model_detection attaches the feature pyramid on
different levels of the backbone.
More precisely, the backbone has for different levels a layer specified
as docking layer. When creating a detection
model, the feature pyramid is attached on the corresponding docking layer.
The pretrained classifiers provided by HALCON have already
specified docking layers. But when you use a self-provided classifier as
backbone, you have to specify them yourself.
You can set backbone_docking_layers as part of the classifier
using the operator set_dl_model_param or the backbone as such
using this operator.
The docking layers are from different levels and therefore the feature maps used in the feature pyramid are of different size. More precisely, in the feature pyramid the feature map lengths are halved with every level. By implication, the input image lengths need to be halved for every level. This means, the network architectures allow changes concerning the image dimensions, but the dimensions 'image_width' and 'image_height' need to be an integer multiple of . Here, is the highest level up to which the feature pyramid is built. This value depends on the attached networks as well as on the docking layers. For the provided classifiers the list below mentions, up to which levels the feature pyramid is built using default settings.
HALCON provides the following pretrained classifiers you can read in as backbone:
This neural network is designed for simple classification tasks. It is characterized by its convolution kernels in the first convolution layers, which are larger than in other networks with comparable classification performance (e.g., 'pretrained_dl_classifier_compact.hdl'). This may be beneficial for feature extraction.
This backbone expects the images to be of the type 'real'.
Additionally, the backbone is designed for certain image properties.
The corresponding values can be retrieved with
get_dl_model_param. Here we list the default values with
which the classifier has been trained:
'image_width': 224
'image_height': 224
'image_num_channels': 3
'image_range_min': -127
'image_range_max': 128
The default feature pyramid built on this backbone goes up to level 4.
This neural network is designed to be memory and runtime efficient.
This backbone expects the images to be of the type 'real'.
Additionally, the backbone is designed for certain image properties.
The corresponding values can be retrieved with
get_dl_model_param. Here we list the default values with
which the classifier has been trained:
'image_width': 224
'image_height': 224
'image_num_channels': 3
'image_range_min': -127
'image_range_max': 128
The default feature pyramid built on this backbone goes up to level 4.
This neural network has more hidden layers than 'pretrained_dl_classifier_compact.hdl' and is therefore assumed to be better suited for more complex tasks. But this comes at the cost of being more time and memory demanding.
This backbone expects the images to be of the type 'real'.
Additionally, the backbone is designed for certain image properties.
The corresponding values can be retrieved with
get_dl_model_param. Here we list the default values with
which the classifier has been trained:
'image_width': 224
'image_height': 224
'image_num_channels': 3
'image_range_min': -127
'image_range_max': 128
The default feature pyramid built on this backbone goes up to level 5.
As the network 'pretrained_dl_classifier_enhanced.hdl', this network is suited for more complex tasks. But its structure differs, bringing the advantage of making the training more stable and being internally more robust.
This backbone expects the images to be of the type 'real'.
Additionally, the backbone is designed for certain image properties.
The corresponding values can be retrieved with
get_dl_model_param. Here we list the default values with
which the classifier has been trained:
'image_width': 224
'image_height': 224
'image_num_channels': 3
'image_range_min': -127
'image_range_max': 128
The default feature pyramid built on this backbone goes up to level 5.
Parameters you can set for your model when creating it using
create_dl_model_detection (see get_dl_model_param for
explanations):
'anchor_angles'
'anchor_aspect_ratios' (legacy: 'aspect_ratios')
'anchor_num_subscales' (legacy: 'num_subscales')
'backbone_docking_layers'
'bbox_heads_weight', 'class_heads_weight'
'capacity'
'class_ids'
'class_ids_no_orientation'
'class_weights'
'freeze_backbone_level'
'ignore_direction'
'image_height', 'image_width'
'image_num_channels'
'instance_type'
'max_level', 'min_level'
'max_num_detections'
'max_overlap'
'max_overlap_class_agnostic'
'min_confidence'
'optimize_for_inference'
To successfully set 'gpu' parameters, cuDNN and cuBLAS are
required, i.e., to set the parameter GenParamName
'runtime' to 'gpu'.
For further details, please refer to the “Installation Guide”,
paragraph “Requirements for Deep Learning”.
This operator returns a handle. Note that the state of an instance of this handle type may be changed by specific operators even though the handle is used as an input parameter by those operators.
Backbone (input_control) filename.read → (string)
Deep learning classifier, used as backbone network.
Default value: 'pretrained_dl_classifier_compact.hdl'
List of values: 'pretrained_dl_classifier_alexnet.hdl', 'pretrained_dl_classifier_compact.hdl', 'pretrained_dl_classifier_enhanced.hdl', 'pretrained_dl_classifier_resnet50.hdl'
File extension: .hdl
NumClasses (input_control) integer → (integer)
Number of classes.
Default value: 3
DLModelDetectionParam (input_control) dict → (handle)
Parameters for the object detection model.
Default value: []
DLModelHandle (output_control) dl_model → (handle)
Deep learning model for object detection.
If the parameters are valid, the operator create_dl_model_detection
returns the value 2 (H_MSG_TRUE). If necessary, an exception is raised.
set_dl_model_param,
get_dl_model_param,
apply_dl_model,
train_dl_model_batch
Deep Learning Training