Method for training an artificial neural network
A method for training an artificial neural network uses training data that include first image data of a first image and second image data of a second image of an infrastructure. The first image includes a first feature, and the second image includes a second feature corresponding to the first image. The training data include a relative desired translation and a relative desired rotation between the first feature and the second feature. The training includes extracting the first feature from the first image and extracting the second feature from the second image using the artificial neural network. The extracted first feature is represented by first feature data having a first volume of data. The extracted second feature is represented by second feature data having a second volume of data. The training further includes ascertaining a relative translation and a relative rotation between the extracted first feature and the extracted second.
1. A method for training an artificial neural network comprising:
receiving training data signals comprising training data for training the artificial neural network, the training data comprises first image data of a first image of an infrastructure and second image data of a second image of the infrastructure, the first image comprises a first feature, and the second image comprises a second feature corresponding to the first image, the training data further comprises a relative target translation and a relative target rotation between the first feature and the second feature;
training the artificial neural network based on the training data, the training comprises:
extracting the first feature from the first image using the artificial neural network, and
extracting the second feature from the second image using the artificial neural network, the extracted first feature is represented by first feature data having a first data volume, and the extracted second feature is represented by second feature data having a second data volume;
ascertaining a relative translation and a relative rotation between the extracted first feature and the extracted second feature using the artificial neural network; and
training the artificial neural network until a loss function that is dependent on a pose of the feature and on the first data volume and/or the second data volume is at a minimum or is less than or less than or equal to a predefined loss function threshold value.
2. The method according to claim 1 , wherein:
the artificial neural network comprises a CNN (convolutional neural network (“CNN”) and a fully connected network (“FCN”),
an output of the CNN is connected to an input of the FCN, and
the first image data and the second image data are provided to an input of the CNN.
3. The method according to claim 1 , wherein the extracted first and second features are quantized such that the first feature data and the second feature data are integers.
4. The method according to claim 1 , wherein:
the loss function is defined by the following mathematical equation: L total =L pose +L data volume ,
L total is the loss function,
L pose is another loss function related to the pose of the feature, and
L data volume is a further loss function related to the first data volume of the first feature data and/or related to the second data volume of the second feature data.
5. The method according to claim 4 , wherein:
L data volume =γ [H ( z )],
γ is a first parameter,
is an expected value,
z is the first and/or second feature data, and
H(z) is the first and/or second data volume.
6. The method according to claim 4 , wherein:
L pose = [|Δt target −Δt determined | 2 ]+βI [|Δq target −Δq determined | 2 ],
is an expected value,
Δt target is the relative target translation,
Δt determined is the relative translation,
β is a hyperparameter,
Δq target is the relative target rotation, and
Δq determined is the relative rotation.
7. The method according to claim 4 , wherein:
L pose = [|Δt target −Δt determined | 1 ]e −s t +s t + [|Δq target −Δq determined | 1 ]e −s q +s q ,
is an expected value,
Δt target is the relative target translation,
Δt determined is the relative translation,
Δq target is the relative target rotation,
Δq determined is the relative rotation,
s t is a second parameter, and
s q is a third parameter.
8. An artificial neural network trained by the method according to claim 1 .
9. A method for extracting a feature from an image of surroundings of a motor vehicle, comprising:
receiving image data signals which represent image data of an image of the surroundings of the motor vehicle;
extracting the feature from the image using the artificial neural network according to claim 8 ; and
outputting feature signals which represent the extracted feature.
10. A device configured to perform the method according to claim 1 .
11. The method according to claim 1 , wherein a computer program includes instructions that, when the computer program is executed by a computer, prompt the computer to carry out the method.
12. The method according to claim 11 , wherein the computer program is stored on a non-transitory machine-readable storage medium.