Generating images for training and testing machine learning models

Abstract

Method (100) for generating example images (4) for training, validating and / or testing an image-processing machine learning model (1), comprising the steps: • a source image (2) recorded with at least one sensor or synthetically generated is provided (110); • a parameterized sensor model (3) is provided (120) which models the influence of at least one effect effective when recording an image with a real sensor on the recorded image, taking into account the physical mode of action of this effect; • the parameterized sensor model (3) is applied (130) to the source image (2) so that an example image (4) is created; • the example image (4) is processed (140) with the machine learning model (1) to produce an output (5); • this output (5) is evaluated (150) with a cost function (6) that provides a measure of the performance of the machine learning model (1); and • the parameters (3a) of the parameterized sensor model (3) are optimized (160) to the goal that, based on these parameters (3a) of the parameterized sensor model, example images (4) are created on which the performance of the machine learning model (1) is worse.

Description

[0001] The present invention relates to the generation and processing of images that can be used as example images for training and testing image-evaluating machine learning models, such as image classifiers. These image-evaluating machine learning models can then be used, for example, in controlling vehicles or robots. State of the art

[0002] The at least partially automated driving of vehicles or robots on company premises or in public transport requires that the vehicle's or robot's surroundings be continuously monitored by measurement technology, and that the measurement data thus obtained be evaluated with a view to planning the vehicle's or robot's further behavior. Machine learning models are used in particular for this evaluation. If such a model is trained with a finite set of training examples with sufficient variability and delivers accurate results for these training examples, it is assumed, due to the generalization power of machine learning models, that the machine learning model is also capable of accurately evaluating unseen data.

[0003] In supervised training, the machine learning model is presented with training examples that are labeled with an expected target output. The deviation of the actual output generated from the training example from the target output is used as feedback for optimizing the machine learning model. Manually labeling a large number of real-world images is time-consuming and expensive. Therefore, there is a need to use synthetically generated training examples for which the label is known in advance and does not have to be determined manually. Disclosure of the invention

[0004] The invention provides a method for generating example images for training, validating, and / or testing an image-processing machine learning model. During training, parameters that characterize the behavior of the machine learning model are optimized. During validation, this optimization can be repeated with several different hyperparameters, which, for example, determine the topology and / or size of the machine learning model or the optimization strategy. In each case, it can be tested how changing the hyperparameters affects the performance of the machine learning model. The finally trained machine learning model can then be tested with yet different data and released for use.

[0005] The method starts with a source image that was captured with at least one sensor or synthetically generated. This source image has semantic content, such as a traffic situation with various traffic-relevant objects. The goal of the method is to process this source image in such a way that it still has essentially the same semantic content, but is particularly well-suited as a sample image for the training process of a machine learning model.

[0006] For this purpose, a parameterized sensor model is provided that models the influence of at least one effect, e.g., an optical effect, on the captured image, taking into account the physical mechanism of this effect. This parameterized sensor model is applied to the source image, creating a sample image. The source image is thus converted into the sample image according to one or more functions stored in the sensor model that describe the effect on the image signal and depend on the parameters of the sensor model.

[0007] The resulting sample image is processed using the machine learning model to produce an output. This output is evaluated using a cost function that provides a measure of the machine learning model's performance. The cost function can, for example, measure the correspondence between the machine learning model's output and a target output known from the semantic content of the sample image.

[0008] The parameters of the parameterized sensor model are optimized to maximize the cost function of the machine learning model. This results in sample images based on these parameters of the parameterized sensor model, on which the machine learning model performs poorly. As explained later, these sample images can be used in particular to confront the machine learning model with more difficult training examples. This can ultimately improve both the generalization ability of the machine learning model and its robustness.

[0009] In this way, specific example images (degraded or with sensor effects) are generated, the occurrence of which can also be expected when the machine learning model is later used with real-life images, and which, at the same time, pose particular difficulties for the machine learning model during evaluation. This can be used, for example, to specifically direct further training of the machine learning model towards sensor effects that it still has deficits in dealing with. Instead of simply increasing the number of example images and thus improving average performance, weaknesses are specifically identified and incorporated into the training process. This is particularly important for the use of machine learning models for image analysis in the context of at least partially automated driving of vehicles or robots. In real-world applications, it is less important that the average evaluation result is good.Much more important is that a certain minimum performance is never undercut. For example, a pedestrian or other traffic-relevant object must never be classified as a freely passable area. This is not yet fully guaranteed by the method proposed here, but the method works toward this goal.

[0010] In particular, the method proposed here can bring synthetically generated source images closer to the domain and / or distribution of the images that a real sensor system will later provide to the machine learning model. Synthetically generated images, in particular, are often too perfect in the sense that they are free of artifacts, noise, and other impairments that normally occur during image acquisition. Thus, real-world and synthetically generated images differ not only in features easily recognizable to the human eye, such as a different composition or spatial distribution of object instances, but also in subtle features that are not recognizable to the human eye, such as patterns in textures in images that are not recognizable to the human eye.A machine learning model with a large number of trainable parameters can learn specific patterns found in the training examples, including patterns resulting from synthetic generation. It can use all of these patterns to solve the assigned task. In later real-world use of the machine learning model, patterns resulting from synthetic generation, in particular, may suddenly be missing. This is one of several possible reasons why the performance of machine learning models trained on synthetically generated example images is often worse than that of machine learning models trained on real-world example images.

[0011] To improve training for generating synthetic data and / or training with the generated synthetic data, in a particularly advantageous embodiment, the parameters of the parameterized sensor model are additionally optimized to ensure that the sample images are realistic according to a predefined criterion. A completely obscured image, for example, does not meet this criterion, so training is directed toward generating a sample image that is impaired but still barely recognizable.

[0012] The specified criterion can be, for example, • a proximity and / or similarity to the source image, and / or to a domain and / or distribution of source images, and / or • a proximity and / or similarity to at least one existing example image, and / or to a domain and / or distribution of existing example images; and / or • include at least one statistical measure of the pixel values ​​of the sample image. This criterion can be weighted in any way relative to the goal of degrading the performance of the trained machine learning model.

[0013] The use of a parameterized sensor model that explicitly models the occurring effects has the advantage that the range of influences that can be generated with it is somewhat restricted from the outset towards realistically expected influences. This makes it difficult for the optimization of the sensor model's parameters to simply take the path of least resistance and generate a sample image in which the semantic content of the source image is completely obscured, such as a sample image filled with a solid color.

[0014] Explicitly modeling the effects also offers the advantage of leveraging specific, existing prior knowledge about the physical mechanism of these effects. In principle, it would also be possible to model the effects generally using a trainable machine learning model. However, this would then be a "black box" that requires no prior knowledge of the effects, but on the other hand, it also does not allow for the consideration of this prior knowledge, even if it exists.

[0015] Nevertheless, it cannot be ruled out that the sensor model is also implemented in whole or in part as a machine learning model, as long as at least the use of a physical mechanism of action of an effect is still recognizable.

[0016] Which optimization method can be used depends, among other things, on the specific sensor model. If the sensor model only subjects the source image to differentiable operations, gradient-based optimization methods can also be used. For example, values ​​of a cost function can be backpropagated through the sensor model to changes in the parameters. Otherwise, any gradient-free method can be used, such as Bayesian optimization or random search using a fast surrogate model for the image-evaluating machine learning model.

[0017] For example, different sensor models can also be combined with each other, for example by calculating the example images generated from one and the same source image.

[0018] In a particularly advantageous embodiment, with the sensor model parameters fixed, those parameters that characterize the behavior of the machine learning model are optimized with the goal of improving the machine learning model's performance on a set of sample images. This counteracts the deterioration in the machine learning model's performance that was initially caused by optimizing the sensor model's parameters. Ideally, the initial deterioration is overcompensated by the machine learning model learning to draw accurate conclusions even from the less recognizable sample images.

[0019] Since the sensor model cannot model the imaging process with absolute accuracy, and in particular, the strength of any effect that occurs in real-world operation (such as rain or dirt on the camera lens) is difficult to predict, it is to be expected that the domain and / or distribution of the sample images will not be completely identical to the domain and / or distribution of the real images presented during operation of the machine learning model. However, if the deviations are such that the sample images show similar effects to the real images, only to a greater extent, this is not necessarily detrimental to the further training of the machine learning model.

[0020] Advantageously, the parameters of the sensor model, on the one hand, and the parameters of the machine learning model, on the other, are continuously optimized alternately. The switch is triggered in response to a predefined criterion being met regarding the amount of data processed, the computing time consumed, and / or the convergence of the optimization. In this way, the total time available for training can be distributed in such a way that the best end result in terms of the performance of the machine learning model is achieved.

[0021] In another particularly advantageous embodiment, during the optimization of the sensor model parameters, the relative weighting between the goals of a degraded performance of the machine learning model on the one hand and the most realistic example images on the other hand is shifted towards a greater weighting of the degraded performance of the machine learning model. In particular, training can initially begin with pure training to generate the most realistic example images possible, before gradually taking into account the degradation in the performance of the machine learning model within the framework of an "annealing schedule." This initially guides the training of the sensor model in the particularly important direction of ensuring that the resulting example images are realistic, before seeking an alternative optimum.This particularly reinforces the tendency to progressively increase the difficulty level of training the machine learning model.

[0022] Alternatively, or in combination with this, sample images can be used during the optimization of the machine learning model's parameters, which are increasingly influenced by the effects of the parameterized sensor model. This ensures a good initialization of the machine learning model with regard to its main task of image analysis before the machine learning model devotes itself to the additional task of deciphering sample images despite the interference effects that occur.

[0023] In the parameterized sensor model, any sensor effects can be modeled and combined with each other. Examples of such effects include: • the influence of a windscreen of a vehicle located in the beam path from a light source to the sensor used for image capture, which can be found, for example, in ◯ the effect of wiper blades, ◯ foreign substances such as streaks, dirt, fog or rain and ◯ aberration and filtering effects of the glass; • Motion blur, for example due to relative movement between the sensor and the observed scene; • the influence of at least one optical element arranged in the beam path in front of the sensor in a device that contains the image capture device for the sensor, such as lens flare, scattered light, distortions, optical aberration and filter effects, focus effects and aperture effects (such as depth of field, aberration or vignetting); • the influence of at least one characteristic of the sensor used for image capture, such as exposure, resolution, arrangement of subpixels for the primary colors (“color sampling”), noise behavior, quantization effects, or the stitching of the output image from multiple exposures; and • the influence of at least one post-processing operation on the raw data supplied by the sensor used to capture the image, such as white balance, correction of interpolation and mosaic effects (“debayering” of colour values), rectification, colour compression, denoising, tonal correction or sharpening.

[0024] Each of these effects, which can also be referred to as "image augmentation," has its own parameters that determine the strength and properties of the respective effect. The final camera model, i.e., the applied "effect chain," is then obtained by combining the individual effects. In particular, the strengths and order of the effects can be varied.

[0025] In another particularly advantageous embodiment, a source image is selected that has a higher dynamic range than the dynamic range expected by the machine learning model for input images. The example image is only converted to the dynamic range expected by the machine learning model after the parameterized sensor model has been applied. In this way, the effects of the sensor model can be calculated with high accuracy. Rounding and other quantization errors are avoided, particularly when combining multiple effects. The dynamic range can be specified, for example, in the number of bits to be used per pixel and color channel, for example, 8 bits, 16 bits, or 32 bits.

[0026] In a further advantageous embodiment, at least one parameter of the sensor model is varied using a sample drawn from a random distribution. This allows a large number of sample images with extensive variability to be generated, which ultimately improves the generalization capability of the machine learning model to be trained on the data.

[0027] The machine learning model can be designed, for example, to use an input image • to recognize object instances and / or object types of these object instances; and / or • to classify the image or parts thereof; and / or • to segment the image semantically; and / or • to determine a pose of at least one object, such as a pedestrian.

[0028] These tasks are often performed by machine learning models when analyzing images of traffic situations in particular.

[0029] Therefore, in another particularly advantageous embodiment, the sensor model models at least one effect that is effective in a given sensor module carried by a vehicle or intended for installation in a vehicle. These sensor modules have specific properties with regard to image quality. The images can therefore have different properties than, for example, synthetically generated images. At the same time, it would be very complex to have to physically recreate all possible operating situations of such a sensor module, such as contamination effects.

[0030] In particular, a sensor model for a specific camera arrangement can be created before this camera arrangement has actually been built. The performance ultimately achieved by the machine learning model can then be used to determine how many example images and what type of example images are needed to successfully train an analysis of images from precisely this camera arrangement. Furthermore, conclusions can also be drawn as to whether any changes to the camera arrangement might significantly improve the performance of the image analysis with the same training effort. To this end, the optimization can be repeated, for example, starting with a modified candidate camera arrangement. The result could then be that replacing a camera lens with a higher-quality lens that costs only slightly more promises a disproportionate gain in the analysis performance.Assuming sufficient computing power, it is possible, for example, to successively determine which change has the best cost-benefit ratio.

[0031] As previously explained, the method proposed here aims to improve a given machine learning model for image analysis so that it delivers better and more reliable results. Therefore, in another particularly advantageous embodiment, additional input images are fed to the machine learning model, which was trained with the example images generated according to the method proposed here. These additional input images were captured with at least one sensor. A control signal is determined from the output generated by the machine learning model. A vehicle, a driver assistance system, a robot, a quality control system, an area monitoring system, and / or a medical imaging system is controlled with the control signal.In this context, the improved training has the effect that the reaction of the respective controlled technical system to the control signal is more likely to be appropriate to the situation embodied in the input images.

[0032] The method can, in particular, be fully or partially computer-implemented. Therefore, the invention also relates to a computer program with machine-readable instructions that, when executed on one or more computers and / or compute instances, cause the computer(s) and / or compute instances to execute the described method. In this sense, control units for vehicles and embedded systems for technical devices that are also capable of executing machine-readable instructions are also to be regarded as computers. Compute instances can, for example, be virtual machines, containers, or serverless execution environments, which can be provided in a cloud, in particular.

[0033] The invention also relates to a machine-readable data carrier and / or a downloadable product containing the computer program. A downloadable product is a digital product that can be transmitted over a data network, i.e., downloaded by a user of the data network, and which can be offered for immediate download, for example, in an online shop.

[0034] Furthermore, one or more computers and / or compute instances may be equipped with the computer program, the machine-readable data carrier or the download product.

[0035] Further measures improving the invention are presented in more detail below together with the description of the preferred embodiments of the invention with reference to figures. Examples of implementation

[0036] It shows: Fig. 1 embodiment of the method 100 for generating example images 4 for training, validation and / or testing of an image processing machine learning model 1; Fig. 2 Illustration of an optimization of the parameterizable sensor model to generate example images 4 with sensor effects on both the degradation of the performance of the machine learning model 1 and on realistic example images 4; Fig. 3 Example processing of a source image 2 to an example image 4 with a variety of effects.

[0037] Fig. 1 is a schematic flow diagram of an embodiment of the method 100 for generating example images 4 for training, validation and / or testing of an image-processing machine learning model 1.

[0038] This image-processing machine learning model 1 can be designed, for example, according to block 105, to use an input image • to recognize object instances and / or object types of these object instances; and / or • to classify the image or parts thereof; and / or • to segment the image semantically; and / or • determine a pose of at least one object.

[0039] In step 110, a source image 2 recorded with at least one sensor or synthetically generated is provided.

[0040] According to block 111, a source image 2 may be selected that has a higher dynamic range than the dynamic range expected by the machine learning model 1 for input images.

[0041] In step 120, a parameterized sensor model 3 is provided. This sensor model 3 models the influence of at least one effect effective when capturing an image with a real sensor on the captured image, taking into account the physical mode of action of this effect.

[0042] According to block 121, a parameterized sensor model can be selected that models at least one of the following effects: • the influence of a vehicle windscreen located in the beam path from a light source to the sensor used to capture the image; • Motion blur; • the influence of at least one optical element arranged in the beam path in front of the sensor in a device containing the image recording device for the sensor; • the influence of at least one characteristic of the sensor used for image acquisition; and • the influence of at least one post-processing of the raw data provided by the sensor used for image acquisition.

[0043] In step 130, the parameterized sensor model 3 is applied to the source image 2 according to the parameters 3a that characterize its behavior. This creates an example image 4.

[0044] According to block 131, the example image 4 can only be converted to the dynamic range expected by the machine learning model 1 after applying the parameterized sensor model 3, insofar as it was previously processed with increased dynamic range according to block 111.

[0045] According to block 132, at least one parameter 3a of the sensor model 3 can be varied using a sample drawn from a random distribution.

[0046] In step 140, the example image 4 is processed with the machine learning model 1 to produce an output 5.

[0047] In step 150, this output 5 is evaluated with a cost function 6, which provides a measure of the performance of the machine learning model 1. The result is an evaluation 6a.

[0048] In step 160, the parameters 3a of the parameterized sensor model 3 are optimized to the goal of degrading the performance of the machine learning model 1 on the additional example images 4 generated based on these parameters 3a. This means that example images 4 are specifically generated with which the machine learning model has problems and does not currently deliver good results.

[0049] According to block 161, the parameters 3a of the parameterized sensor model 3 can be additionally optimized to the goal that the example images 4 are realistic according to a predetermined criterion.

[0050] According to block 161a, this predetermined criterion can in particular be, for example, • a proximity and / or similarity to the source image 2, and / or to a domain and / or distribution of source images 2, and / or • a proximity and / or similarity to at least one already existing example image 4, and / or to a domain and / or distribution of already existing example images 4; and / or • contain at least one statistical characteristic of the pixel values ​​of the example image 4.

[0051] According to block 162, in the course of optimizing the parameters 3a of the sensor model 3, a relative weighting between the goals of a degraded performance of the machine learning model 1 on the one hand and the most realistic example images 4 on the other hand can be shifted towards a stronger weighting of the degraded performance of the machine learning model 1.

[0052] For this purpose, for example, a total cost function (loss function) L be set up, the • the cost function 6 for evaluating the performance of the machine learning model 1 as a first contribution Ltask as well as • the measure of how realistic the example images 4 are, as a second contribution Limage The two contributions can, for example, be L=(1−α)⋅Ltask+α⋅Limage where the relative weight α ∈ [0,1] can be varied, for example, within the framework of an arbitrary “annealing schedule”.

[0053] The state of the parameters 3a after optimization is designated by the reference symbol 3a* and defines a “finished” state 3* of the sensor model 3.

[0054] In the Fig. In the example shown in Figure 1, in step 170, with parameters 3a of the parameterized sensor model 3 fixed, parameters 1a that characterize the behavior of the machine learning model 1 are optimized with the goal of improving the performance of the machine learning model 1 on a set of example images 4. In particular, the processing of those example images 4 with which the machine learning model 1 has previously had difficulties can be improved.

[0055] In particular, for example, according to block 171, example images 4 can be used in the course of this optimization, which are characterized with increasing intensity by effects of the parameterized sensor model 3.

[0056] According to blocks 180a and 180b, in particular, for example, the parameters 3a of the sensor model 3 on the one hand and the parameters 1a of the machine learning model 1 on the other hand can be continuously optimized alternately. In particular, after an optimization of the machine learning model 1 in step 170, new example images 4 can be generated, which in turn represent a new challenge for the optimized machine learning model 1. In a further iteration of the training, the machine learning model 1 can in turn learn to master this new challenge. In this case, the change can be triggered in response to a predetermined criterion regarding the amount of processed data, the computing time consumed, and / or the convergence of the optimization being met.

[0057] The state of the parameters 1a after optimization is denoted by the reference symbol 1a* and defines the “fully trained” state 1* of the machine learning model 1.

[0058] In the Fig. In the example shown in Figure 1, in step 190, further input images 7, which were recorded with at least one sensor 8, are fed to the fully trained machine learning model 1*. A control signal 200a is determined from the outputs 5 generated thereby. In step 210, a vehicle 50, a driver assistance system 51, a robot 60, a quality control system 70, an area monitoring system 80, and / or a medical imaging system 90 are controlled with this control signal 200a.

[0059] Fig. Figure 2 illustrates how the parameters 3a of the sensor model 3 can be optimized to avoid deterioration of the performance of the machine learning model 3 on the one hand and to generate realistic example images 4 on the other hand.

[0060] In the Fig. In the example shown in Figure 2, the sensor model 3 comprises a parameter generation unit 3b for generating sensor effect parameters 3a and an effect sampling unit 3c for compiling specific sensor effects based on these parameters 3a. The parameter generation unit 3b also receives a noise vector η drawn from a random distribution as input.

[0061] Sensor model 3 is applied to source image 2 according to step 130 of method 100, resulting in a sample image 4. From this sample image 4, an output 5 is generated in step 150 of method 100 using machine learning model 1, which is then assigned a rating 6a by cost function 6. This rating 6a is fed back to parameter generation unit 3b to generate new sensor effect parameters 3a.

[0062] In the Fig. In the example shown in Figure 2, the sample image 4 is additionally compared with a source image 2' that is different in content but similar in style and is considered realistic in the context of the current application. The result of this comparison is also fed back to the parameter generation unit 3b and influences the generation of the new sensor effect parameters 3a.

[0063] Fig. Figure 3 illustrates the path from a source image 2 through the application of a multitude of effects 31a-31l of a sensor model 3 to the example image 4. The effects 31a-31l are applied one after the other, and if their order is changed, the resulting example image 4 will usually also change. In the Fig. In the example shown in Figure 3, the following effects are applied: • directional blur as effect 31a; • Drops (e.g. rain) in the beam path as effect 31b; • Scattered light effects as effect 31c; • Blur and chromatic aberration as effect 31d; • a scaling of the intensity up to saturation as effect 31e; • Color properties as effect 31f; • Poisson noise and Gaussian noise as effect 31g; • a quantization of the signal as effect 31h; • a white balance as effect 31i; • a color reconstruction as effect 31j; • a tonal correction as effect 31k; and • Sharpen as effect 311.

Claims

[1] Method (100) for generating example images (4) for training, validating and / or testing an image processing machine learning model (1), comprising the steps: • a source image (2) recorded with at least one sensor or synthetically generated is provided (110); • a parameterized sensor model (3) is provided (120) which models the influence of at least one effect effective when capturing an image with a real sensor on the captured image, taking into account the physical mode of action of this effect; • the parameterized sensor model (3) is applied (130) to the source image (2) so that an example image (4) is created; • the example image (4) is processed (140) with the machine learning model (1) to produce an output (5); • this output (5) is evaluated (150) with a cost function (6) that provides a measure of the performance of the machine learning model (1); and • the parameters (3a) of the parameterized sensor model (3) are optimized (160) to the goal that, based on these parameters (3a) of the parameterized sensor model, example images (4) are created on which the performance of the machine learning model (1) is worse. [2] Method (100) according to claim 1, wherein the parameters (3a) of the parameterized sensor model (3) are additionally optimized (161) to the goal that the example images (4) are realistic according to a predetermined criterion. [3] Method (100) according to claim 2, wherein the predetermined criterion • a proximity and / or similarity to the source image (2), and / or to a domain and / or distribution of source images (2), and / or • a proximity and / or similarity to at least one already existing example image (4), and / or to a domain and / or distribution of already existing example images (4); and / or • contains at least one statistical characteristic of the pixel values ​​of the example image (4) (161a). [4] Method (100) according to one of claims 1 to 3, wherein, with fixed parameters (3a) of the parameterized sensor model (3), parameters (1a) that characterize the behavior of the machine learning model (1) are optimized (170) to the goal of improving the performance of the machine learning model (1) on a set of example images (4). [5] Method (100) according to claim 4, wherein • the parameters (3a) of the sensor model (3) on the one hand and the parameters (1a) of the machine learning model (1) on the other hand are continuously optimized alternately (180a, 180b), whereby • the change is triggered in response to a given criterion being met regarding the amount of data processed, the computing time used, and / or the convergence of the optimization. [6] Method (100) according to one of claims 2 to 5, wherein • in the course of optimising the parameters (3a) of the sensor model (3), a relative weighting between the objectives of a degraded performance of the machine learning model (1) on the one hand and of the most realistic example images (4) on the other hand is shifted towards a greater weighting of the degraded performance of the machine learning model (1) (162), and / or • in the course of optimizing the parameters (1a) of the machine learning model (1), example images (4) are used (171), which are characterized with increasing strength by effects of the parameterized sensor model (3). [7] Method (100) according to one of claims 1 to 6, wherein a parameterized sensor model (3) is selected (121) which models at least one of the following effects: • the influence of a vehicle windscreen located in the beam path from a light source to the sensor used to capture the image; • Motion blur; • the influence of at least one optical element arranged in the beam path in front of the sensor in a device containing the image recording device for the sensor; • the influence of at least one characteristic of the sensor used for image acquisition; and • the influence of at least one post-processing of the raw data provided by the sensor used for image acquisition. [8] Method (100) according to one of claims 1 to 7, wherein • a source image (2) is selected (111) which has a higher dynamic range than the dynamic range expected by the machine learning model (1) for input images; and • the example image (4) is only converted to the dynamic range expected by the machine learning model (1) after the parameterized sensor model (3) has been applied (131). [9] Method (100) according to one of claims 1 to 8, wherein at least one parameter (3a) of the sensor model (3) is varied (132) based on a sample drawn from a random distribution. [10] Method (100) according to one of claims 1 to 9, wherein a machine learning model (1) is selected (105) which is designed to, based on an input image • to recognize object instances and / or object types of these object instances; and / or • to classify the image or parts thereof; and / or • to segment the image semantically; and / or • determine a pose of at least one object. [11] Method (100) according to one of claims 1 to 10, wherein the sensor model (3) models at least one effect that is effective in a predetermined sensor module carried by a vehicle or intended for installation in a vehicle. [12] Method according to one of claims 1 to 11, wherein • the machine learning model (1 *) trained with the example images (4) generated according to the method (100) is supplied (190) with further input images (7) which were recorded with at least one sensor (8); • a control signal (200a) is determined (200) from the output (5) generated by the machine learning model (1*) and • a vehicle (50), a driver assistance system (51), a robot (60), a quality control system (70), a system (80) for monitoring areas, and / or a medical imaging system (90), with which the control signal (200a) is controlled (210). [13] A computer program comprising machine-readable instructions which, when executed on one or more computers and / or compute instances, cause the computer(s) and / or compute instances to carry out the method (100) according to any one of claims 1 to 12. [14] Machine-readable data carrier and / or download product with the computer program according to claim 13. [15] One or more computers and / or compute instances with the computer program according to claim 13, and / or with the machine-readable data carrier and / or download product according to claim 14.

Images