| CPC G06V 20/56 (2022.01) [G06T 9/00 (2013.01); G06V 10/60 (2022.01); G06V 2201/07 (2022.01)] | 14 Claims |
|
1. A method for generating encoded training data, wherein
a processing unit performs the following steps for generating the encoded training data and for storing them into a data storage during a test drive of a test vehicle that carries a camera that is generating raw image data:
receiving the raw image data from the camera and
operating a first artificial neural network (ANN) that has been trained on raw image training data and that is trained to recognize at least one object on the basis of raw image data, and
performing an image recognition on the basis of at least one data sequence of the received raw image data using the first artificial neural network and thereby generating respective reference recognition data on the basis of a predefined evaluation rule, wherein the reference recognition data describe (i) first image content that has been recognized in the at least one data sequence, (ii) activation values of at least one hidden layer of the first ANN and/or (iii) a value of a predefined uncertainty measure regarding the first image content, and wherein the evaluation rule determines what type of recognition data shall be used, wherein for each data sequence of the raw image data it is tested, if a given video encoder is suitable for encoding the raw image data for generating the encoded training data for a training of another, a second artificial neural network, by performing a testing procedure comprising:
performing the first image recognition on the data sequence using the first artificial neural network;
obtaining the reference recognition data as a result of performing the first image recognition;
encoding the data sequence using the given video encoder that is configured with given preset values, wherein the encoding results in an encoded sequence, and then
decoding the encoded sequence using a corresponding decoder and
performing a second image recognition on the decoded sequence using the first artificial neural network;
obtaining current evaluation data as a result of performing the second image recognition and on the basis of the evaluation rule, wherein the current evaluation data describe second image content that has been recognized in the at least one data sequence, activation values of at least one hidden layer of the first ANN and/or a value of a predefined uncertainty measure regarding the second image content; and
the processing unit verifies, if the current evaluation data and the respective corresponding reference recognition data fulfill a predefined similarity criterion, wherein the similarity criterion comprises the condition that a respective difference value describing a difference between the reference recognition data and the current evaluation data lies within a predefined interval, and
if the current evaluation data and the reference recognition data fulfill the predefined similarity criterion, storing the encoded sequence and/or encoding further received raw image data and storing the encoded further image data as the encoded training data in the data storage;
wherein if the similarity criterion is not fulfilled, the preset values are adapted according to a predefined adaptation rule, wherein the adaptation rule comprises that a stepwise increase or decrease of at least one parameter value of the preset values is performed and/or the adaptation rule comprises that the preset values are adapted as a function of the current evaluation data and wherein after adapting the preset values, the raw image data of the sequence are encoded again and the similarity criterion is verified again.
|