DE102022120240A1 - Method for generating training data through image transformation, method for training a machine learning model and assistance system - Google Patents

Abstract

The invention relates to a method (10) for generating training data (18, 20) for a machine learning model for estimating depth maps from mono camera images (12). A mono camera image (12) and corresponding depth data (14) are provided therein as ground truth data. A transformed camera image (18) with spatially shifted pixel values is generated from the mono camera image (12) using a non-linear, location-dependent transformation, and transformed depth data (20) is generated from the corresponding depth data (14) using the same transformation. A transformed camera image (18) and the corresponding transformed depth data (20) serving as ground truth data are then provided as training data (18, 20). The invention further relates to a method (10) for training a machine learning model for estimating depth maps for mono camera images (12) using appropriately generated training data (18, 20) and a vehicle assistance system with such a trained machine learning model.

Description

The present invention relates to a method for generating training data for training a machine learning model to estimate depth maps from camera images. The invention further relates to a method for appropriately training such a machine learning model and an assistance system with such a machine learning model.

Nowadays, there are efforts to use the methods and possibilities of machine learning in many different areas and applications, as this can in principle achieve advantages over, for example, classic manually programmed algorithms or solutions. However, large amounts of suitable training data are required to generate corresponding machine learning models. Obtaining such training data can be extremely labor-, time- and cost-intensive and therefore represents a significant hurdle for the practical use of corresponding machine learning models.

As an approach to address this problem, the DE 10 2019 103 192 A1 a method for generating training data for an adaptive camera system. A first and a second training data set for different camera types are created by taking a large number of digital images of different scenarios. The images for the second training data set are labeled, while the images for the first training data set are not labeled. A transfer function is then determined depending on information from the first and second training data sets. This transfer function is applied to the second training data set and the second training data set modified in this way is provided as training data for the adaptive camera system. This is intended to enable less cost- and/or time-intensive training, in particular reducing the effort required to generate training data. However, the problem here is that a large number of suitable digital images must first be recorded for the training data, which can involve enormous effort.

An example of a method for training neural networks is in the DE 11 2020 001 897 C5 described. The method there includes receiving map data and locating a dynamic actor in relation to the map data. The method further includes receiving sensor data that was generated by a sensor of the dynamic actor. Based on the map data and localization, ground truth data corresponding to the sensor data is generated. A neural network is then trained using the sensor data and the ground truth data.

The object of the present invention is to enable particularly low-effort and robust generation of depth maps from mono camera images.

This task is solved by the subject matter of the independent patent claims. Further possible embodiments of the invention are disclosed in the subclaims, the description and the figures. Features, advantages and possible configurations set out in the description for one of the subject matter of the independent claims are at least analogous to the features, advantages and possible embodiments of the respective subject matter of the other independent claims and any possible combination of the subject matter of the independent claims, if applicable in conjunction with one or more of the subclaims.

A method according to the invention is used to generate training data for training a machine learning model to estimate distance or depth maps from mono camera images. Such a mono camera image can be a single image that does not contain any direct depth information, i.e. no distance information of the object shown from the image recording device used, i.e. a camera. This can be seen in particular in contrast to a stereo image, from which the distance of displayed objects, i.e. corresponding distance or depth information, can be taken or directly determined objectively and clearly. The method according to the invention can be used in particular for vehicle technology, so that the monocamera image can be recorded in particular by means of a camera of a motor vehicle provided for detecting the surroundings. Likewise, the method according to the invention can in principle be used for other technical areas or applications.

A depth map in the present sense can be a data set which, in particular, can be of the same dimensions as the respective mono camera image and, for each point or pixel, a distance of the object displayed there from the camera by which the respective mono camera image was recorded was, indicates.

The machine learning model can in particular be an artificial neural network, in principle but be or include any machine learning device or structure.

In a method step of the method according to the invention, at least one mono camera image of a scene, for example a single mono camera image or a sequence or series of several mono camera images, and distance or depth data corresponding to the same scene are provided as ground truth data. These depth data indicate distances to individual areas, points or pixels between an object in the scene that is located there, i.e. in particular also shown at the corresponding location in the mono camera image, and a recording device. The depth data can have been recorded using this recording device. Likewise, the recording device may have been used to record the mono camera image or may include the corresponding camera. The camera and the recording device for the depth data can be connected or arranged in particular in a fixed manner or in a fixed position relative to one another. The mono camera image and the associated depth data may have been recorded at least substantially simultaneously.

In order to compensate for or eliminate any differences or deviations between the coordinate systems of the mono camera image and the depth data, the mono camera image and the depth data can, for example, be brought or projected into a common coordinate system. For example, the depth data can be projected or transformed into the coordinate system of the mono camera image or vice versa. This can ensure that the pixels for which the depth data indicate distances can or are uniquely assigned or superimposed on the corresponding pixels of the mono camera image, i.e. those depicting or relating to the same object.

In a further method step of the method according to the invention, at least one transformed monocamera image with spatially shifted image or pixel values compared to the underlying original monocamera image is generated from the monocamera image at least by means of a non-linear, location-dependent transformation. This transformed mono camera image is also referred to here as a transformed camera image to make it easier to distinguish it from the original mono camera image. The transformation can shift pixels or corresponding image or pixel values, such as color values, intensities, brightnesses and/or the like, in at least one direction or dimension of the camera image. For a 2D mono camera image with x and y pixel coordinates, approximately in the x and/or y direction.

If applying the transformation to the mono camera image for a pixel or pixel values of the mono camera image would lead to a new position outside the mono camera image, i.e. outside of its original boundaries or coordinates, the shift for this pixel or these pixel values can, for example, be automatically reduced, to avoid this. Likewise, the transformation or a shift derived from it can be generated from the outset with the boundary condition that it must not lead to shifts outside the limits of the original mono camera image. For example, it can be provided that the pixels or pixel values at the edge of the mono camera image are or may only be moved along the edge, but not beyond it. Likewise, for example, pixels or pixel values that would be shifted beyond the edge can be positioned - analogous to a cyclic permutation - on the mirrored side or location of the mono camera image or the transformed camera image. Likewise, for example, by moving pixels or pixel values beyond the edge, free or missing areas can be automatically filled in a predetermined manner, for example depending on the surrounding pixels or pixel values. Ultimately, it can be ensured that the transformed camera image has the same overall shape and dimensions as the underlying mono camera image.

In a further method step of the method according to the invention, correspondingly transformed depth data is generated from the corresponding depth data using the same transformation that is also used to generate the transformed camera image. The same conditions or mechanisms described in connection with generating the transformed camera image can be applied to ensure that the transformed depth data has the same shape and dimension as the originally provided depth data.

The transformed depth data and the transformed camera image can therefore contain at least essentially the same data as the originally provided depth data and the originally provided mono camera image, although these data or their values can be changed spatially, i.e. with regard to their position or arrangement in the corresponding coordinate system. For example, in the original mono camera image, it is a rectangular object shown, the same object can be shown in the transformed camera image, but have a different shape or contour. For example, the object in the transformed camera image can then no longer be rectangular and/or have curved or wavy lines or the like instead of straight edges or outer contours. If, for example, the original depth data contains a distance information for a point or pixel lying on the outer contour of this object, the transformed depth data can contain a distance information that is shifted in the same way relative to the original depth data with regard to its pixel or data coordinates, i.e. transformed, like the corresponding pixel in the transformed camera image relative to the original mono camera image. The corresponding distance information in the transformed depth data therefore continues to indicate the distance to the corresponding point on the outer contour of the same object, which, however, is located at a different location within the transformed camera image than in the original mono camera image.

In a further method step of the method according to the invention, a camera image transformed in this way and the corresponding transformed depth data serving as ground truth data for this transformed camera image are provided as training data or at least as part of the training data. For this purpose, the training data can, for example, be stored in a predetermined data memory, output via a predetermined interface or fed directly to the machine learning model to be trained as input data, or the like.

The method according to the invention therefore allows computer-aided generation of varied, i.e. effectively new, training data from provided training data. For example, new training data can be generated from mono camera images recorded as training data and associated depth data.

Likewise, the method according to the invention can be used iteratively in order to generate transformed or varied new training data of the next generation, i.e. the next iteration step, from new training data already generated in an earlier run of the method according to the invention, which can then be used as a starting point.

The method according to the invention can save effort for manually recording corresponding training data in real environments or a larger set of training data can be provided with a certain amount of such effort. This makes it possible to generate appropriately trained machine learning models more quickly and cost-effectively, which can also be trained more robustly compared to the exclusive use of real training data. The latter can be the case not only because more training data can be used to train the respective machine learning model, but also because the training data generated using the method according to the invention can have properties that can only be recorded in real environments or situations with an enormous amount of time or purely by chance . The training data generated using the method according to the invention can therefore represent corresponding edge cases that can be particularly important for robust training.

It is particularly advantageous that the non-linear transformation proposed here enables much more complex variations of the original mono camera images and depth data provided as a starting point than, for example, conventional global linear transformations, such as a simple rotation or scaling, which are homogeneous and at every point has the same effect. In comparison to such global linear transformations, the changes generated by the non-linear transformation provided according to the invention are less easy to learn as patterns by a machine learning model. This can force the machine learning model to learn to a greater or more robust extent the actual fundamental relationships between camera images or the properties or appearance of the displayed objects and their distances from the camera.

In a possible embodiment of the present invention, lidar measurement data, i.e. measurement or sensor data from a lidar sensor, in particular a lidar sensor for a motor vehicle, are used as depth data. These eyelid measurement data only indicate distances for some of the points or pixels. The eyelid measurement data therefore does not contain distance information for every point or pixel and in this sense can be sparse data. The eyelid measurement data and the respective associated mono camera image can be projected or transformed into a common coordinate system. For the sake of simplicity, for example, the camera coordinate system, the lidar coordinate system or a vehicle coordinate system can be used. Since the lidar measurement data used here as ground truth data for the depth map estimation is sparse, i.e. not dense or not complete, and at least conventional lidar sensors have a specific fixed sampling or measuring point pattern, the use could of eyelid measurement data as ground truth data lead to artifacts in estimated depth maps of a machine learning model trained with it. Such artifacts can be, for example, horizontal stripe patterns that do not correlate with real conditions. Such artifacts can arise because the pixel or measurement point position of the eyelid measurement points, which may be projected into a different coordinate system, i.e. eyelid detections, and the distance to the object imaged or detected there can be correlated with one another.

This can be the case at least under the condition that the camera for recording the monocamera image and the lidar sensor for recording the lidar measurement data are arranged in a fixed position or in a fixed position relative to one another or are connected to one another. This means that the lidar sensor typically only detects or scans the same points or areas relative to the lidar sensor, for example relative to a motor vehicle equipped with it, while a lidar measurement signal is missing in the remaining areas or points. Based on such data, the machine learning model may then not be able to learn the actual distance depending on the mono camera image, but rather, for example, learn that under the condition that eyelid measurement data is present or available at a certain point, there is, for example, a road there.

This problem can be counteracted by the method according to the invention, since the transformation means that the new, varied training data that is also available for training the machine learning model can also contain lidarm measurement data on areas, points or pixels that are not directly scanned or detected by the respective lidar sensor. Thus, the present invention makes it possible to use eyelid measurement data as ground truth data for training a machine learning model to estimate depth maps based on monocamera images and ultimately an artifact-free or artifact-reduced estimation of depth maps using a correspondingly trained machine learning model.

In a further possible embodiment of the present invention, an elastic deformation is used as the transformation. Since such elastic deformations can significantly change the respective monocamera image transformed with it and in particular also in the relative positional relationships between displayed objects, it is a non-trivial finding that such elastic deformations can still be used as a transformation to generate training data for depth map estimation and training data generated in this way can be successfully used to train a corresponding machine learning model. However, according to a finding underlying the present invention, this is exactly the case.

In a further possible embodiment of the present invention, a noise image with a number of pixels corresponding at least to the number of pixels of the respective mono camera image is generated for the transformation. This noise image can in particular have the same dimension as the respective mono camera image. The image or pixel values of the noise image correspond to noise, in particular random or pseudo-random noise. In other words, the image or pixel values of the noise image can, for example, be determined randomly or set using a noise generator. Furthermore, for the transformation from the noise image, displacement data is generated here, in particular by means of a predetermined filter, which indicates or defines a spatial displacement for each pixel. The displacement data can be or comprise, for example, a vector field or two matrices, one of which indicates the displacement in the x direction for each pixel and the other indicates the displacement in the y direction for each pixel. The displacement data can therefore in particular indicate or define the respective spatial displacement in all directions or dimensions, in the case of a 2D mono camera image and noise image with x and y pixel coordinates, i.e. in the x and y directions. The displacement for at least one direction or dimension can also be zero for at least some of the pixels. The displacement data can ultimately define the transformation.

For the transformation, the pixel values of the mono camera image are then shifted according to the shifts specified for the respective pixel in the shift data. This can be done, for example, in a lookup process in which the displacement specified in the displacement data for this pixel is looked up and applied for each pixel of the mono camera image to be transformed.

Through the embodiment of the present invention proposed here, different variations, i.e. different transformed camera images and correspondingly transformed depth data, can be generated in a particularly simple and efficient manner from a single provided monocamera image with associated depth data. Due to the inherently random properties of the noise image, it can be avoided that the training data generated in this way contains a specific fixed pattern that would lead to a false training ning of the respective machine learning model and thus ultimately lead to undesirable artifacts in the depth maps estimated with it.

In order to generate the greatest possible variety of training data from the respective output data, i.e. the originally provided mono camera image and the associated depth data, several or many different noise images can be generated, for example. Likewise, at least one new noise image can be generated for each mono camera image provided as a starting point. Overall, a set of training data can be generated in a particularly simple and efficient manner, which enables particularly effective and robust training of a machine learning model.

In a possible development of the present invention, a predetermined Gaussian filter is applied to the noise image to generate the displacement data. This makes it possible to ensure that, for a particular pixel, the displacements in its neighborhood or surroundings are relatively similar to the displacement intended for the respective pixel. This makes it possible to avoid excessive or chaotic changes or transformations of the respective mono camera image, which could ultimately cause new artifacts in the estimated depth maps. By appropriately parameterizing the Gaussian filter, the neighborhood or environment, i.e. the region, in which shifts of relatively similar size or in the same or a relatively similar direction occur or are made can be determined as required. This allows for flexible adaptation to different requirements as well as optimization to minimize artifacts in the estimated depth maps and maximize the accuracy and robustness of the respective machine learning model.

In a further possible embodiment of the present invention, the shift of the pixel values or a corresponding parameterization of the transformation that determines this is limited to a predetermined maximum value. In other words, the scope or extent of the shifts can be limited in a predetermined manner. This can counteract the emergence of problems or artifacts in the estimated depth maps, i.e. enable the machine learning model to be trained in a correspondingly reliable and robust manner. The predetermined maximum value can, for example, depend on the data used as a starting point, its dimension or image size, details shown therein, the density of the depth data and/or the like. The specified maximum value can be determined, for example, as part of a hyperparameter search or determined experimentally in some other way for the respective application.

In a further possible embodiment of the present invention, a large number of correspondingly different transformed camera images and thus a large number of correspondingly different training data are generated from a specific mono camera image by means of several different or varied transformations. In other words, several different non-linear location-dependent transformations can be applied to one and the same mono camera image and the associated depth data. This means that a larger and more diverse set of training data can be generated overall, which in turn can enable more efficient, effective and robust training of the respective machine learning model. The various transformations can be generated, for example, by random variations, which can avoid, for example, any unintentional introduction of patterns into the training data. If several mono camera images with associated depth data are provided, the same transformations and/or different transformations can be applied to them.

In a further possible embodiment of the present invention, the training data and/or further training data are generated from the mono camera image and the corresponding depth data by applying at least one predetermined global linear transformation. Such a global linear transformation can, for example, be or include a constant, i.e. homogeneous or the same rotation or scaling for all points or pixels. The global linear transformation and the non-linear transformation can each be applied alone, i.e. exclusively, or together, i.e. in combination with one another, to the respective monocamera image and the associated depth data in order to generate the training data. In other words, first training data can be generated by exclusively applying the global linear transformation and separately second training data by exclusively applying the non-linear transformation and / or third training data by combinedly applying both the global linear transformation and the non-linear transformation respectively one and the same mono camera image. This means that more different and diverse training data can be generated from a single mono camera image. This can enable more efficient, effective and robust training of the respective machine learning model with particularly little data provided as a starting point.

The present invention also relates to a method for training a machine learning model, in particular an artificial neural network, for estimating distance or depth maps for mono camera images or from mono camera images. For better distinction, this method can also be referred to as the second method according to the invention, while the method according to the invention described elsewhere for generating training data for a corresponding machine learning model can be referred to as the first method according to the invention.

In a method step of the second method according to the invention, at least one mono camera image and depth data corresponding to it as ground truth data, in particular in the form of eyelid measurement data recorded simultaneously with the respective mono camera image for the same scene, are provided. Training data is generated from this using the first method according to the invention. Using the training data generated in this way, the machine learning model is trained in a further method step of the second method according to the invention, in particular by means of supervised learning. For this purpose, for example, the transformed camera images and the associated transformed depth data can be supplied to the machine learning model as input, in particular at different input points or input layers in the structure of the machine learning model. The second method according to the invention can be used to train the machine learning model, which is also described in connection with the first method according to the invention, or can include the method steps, measures or processes mentioned in connection with the first method according to the invention for training the machine learning model as, if necessary, optional, further method steps.

The present invention also relates to an assistance system for a motor vehicle. The assistance system according to the invention has an interface for acquiring monocamera images and a data processing device. This data processing device includes or implements a machine learning model that is trained according to the second method according to the invention. The assistance system according to the invention is set up to generate depth maps for captured mono camera images. The machine learning model of the assistance system according to the invention can in particular be or correspond to the machine learning model mentioned in connection with the methods according to the invention. Generating the depth maps can be the main purpose or the main function or just a partial purpose or a partial function of the assistance system according to the invention. The assistance system according to the invention can therefore output or provide the generated depth maps to another device, for example via the or an interface, or can further use or further process the generated depth maps themselves, i.e. internally.

The assistance system according to the invention can be used in particular for assisted or at least partially automated operation of a motor vehicle.

The assistance system according to the invention, in particular its data processing device, can, for example, comprise a computer-readable data memory and a processor device coupled thereto. The machine learning model as well as a predetermined operating or computer program, which encodes or implements the functions for processing the captured mono camera images into depth maps using the machine learning model, can be stored in this data memory. This operating or computer program can then be executed using the process device, for example using a microchip, microprocessor or microcontroller or the like, in order to generate the depth maps.

Further features of the invention can emerge from the claims, the figures and the description of the figures. The features and combinations of features mentioned above in the description as well as the features and combinations of features shown below in the description of the figures and/or in the figures alone can be used not only in the combination specified in each case, but also in other combinations or on their own, without the scope of the invention to leave.

• 1 einen beispielhaften schematischen Ablaufplan für ein Verfahren zum Erzeugen von unterschiedlichen Trainingsdaten und zum Trainieren eines Maschinenlernmodells damit;
• 2 eine schematische Darstellung von als Ausgangspunkt dienenden Trainingsdaten aus einem Bild und zugehörigen Lidardaten; und
• 3 eine schematische Darstellung von aus den als Ausgangspunkt dienenden Trainingsdaten erzeugten variierten Trainingsdaten.

The drawing shows in:

• 1 an exemplary schematic flowchart for a method for generating different training data and for training a machine learning model therewith;
• 2 a schematic representation of starting training data from an image and associated lidar data; and
• 3 a schematic representation of varied training data generated from the training data serving as a starting point.

In the figures, identical and functionally identical elements are provided with the same reference numerals. For the sake of clarity, only a representative selection may be explicitly marked.

For various applications, knowledge of the distance in camera images can be required Objects displayed by the recording camera can be useful. If only mono camera images are available that do not contain any direct information about the distances, the distances can be estimated in the form of so-called depth maps from the mono camera images using a machine learning model. A mono camera image can be provided or fed to such a machine learning model as input, from which the machine learning model then generates a depth map for this mono camera image as an output. In this depth map, each pixel value corresponds to the distance between the camera used to capture the mono camera image and the object represented in that pixel. However, such a machine learning model must first be trained. This in turn requires the provision of suitable training data.

This shows 1 an exemplary schematic flowchart 10 for corresponding procedures. In a method step S1a, at least one mono camera image of a specific scene, for example a traffic scene, is recorded or provided. At the same time, in a method step S1b, corresponding eyelid measurement data from the same scene is recorded or provided, which indicates the distances to detected or scanned objects in the scene, i.e. corresponding depth data.

In a method step S2, basic or initial data are generated from the at least one mono camera image and the associated eyelid measurement data. This can include, for example, data preparation, unifying the coordinate systems of the at least one mono camera image and the associated eyelid measurement data and/or the like.

For further illustration shows 2 schematically a corresponding output image 12 and associated lidar data 14 recorded by means of a lidar sensor. The lidar data 14 include a large number of lidar measuring points 16, of which two are explicitly marked here only as an example for the sake of clarity. The eyelid measurement points 16 are present here in a regular, uniform grid or pattern, so they do not indicate a distance for each pixel of the output image 12. However, the lidar data 14 indicate distances actually measured for or in the lidar measuring points 16, i.e. depth data, and can therefore be used as ground truth data regarding the distances for the output image 12.

If the machine learning model were to be trained only with the output data generated in this way, on the one hand a corresponding number of monocamera images and associated eyelid measurement data would have to be recorded and, on the other hand, undesirable behavior of the eyepiece could occur, particularly due to the regularity and sparsity of the eyelid measurement data, which can be seen, for example, in the lidar data 14 machine learning model, which can ultimately lead to artifacts in the depth maps generated by the appropriately trained machine learning model.

To avoid this, further training data for the machine learning model is generated based on the initial data. For this purpose, at least one random noise image can be generated for the at least one mono camera image provided in a method step S3. In a method step S4, displacement data can be generated from this. This may include, for example, applying a Gaussian filter and/or the like. The displacement data defines for each point or pixel a displacement or a corresponding non-linear transformation that is derived from the respective noise image and is therefore at least to a certain extent random. Likewise - if necessary optionally - one or more global linear transformations can be provided, i.e. specified, in a method step S5.

By applying at least the non-linear transformation in a method step S6a to the monocamera images contained in the output data and in a method step S6b to the associated depth data contained in the output data, here the corresponding eyelid measurement data, correspondingly transformed camera images and depth images transformed in the same way are obtained. or eyelid measurement data is generated.

Shows for illustrative purposes 3 a transformed image 18 generated from the original image 12 by applying a transformation. An elastic deformation was used as the transformation here. Next to it shows 3 transformed lidar data 20 generated by applying the same transformation to the lidar data 14. These transformed lidar data 20 contain the same lidar measurement points 16 as the original lidar data 14, but now as lidar measurement points 22 that have been moved to other positions due to the transformation. It can be seen here that due to the transformation originally regular structures of the output image 12 and the originally regular grid pattern of the lidar data 14 are contained in the transformed image 18 and in the transformed lidar data 20 in an irregularly changed form.

Creating such altered data from the underlying source data by applying one or more such transformations nen can also be referred to as augmentation.

The transformed camera images and the associated transformed depth or eyelid measurement data are provided as new or further training data in a method step S7. The training data provided here can also include the underlying initial data.

With the training data provided in this way, in a method step S8, a predetermined machine learning model is trained to estimate, i.e. generate, associated depth maps as output from mono camera images provided as input.

A machine learning model trained in this way can then be used, for example, in an assistance system for a motor vehicle or the like.

Overall, the examples described show how artifact-free monitored monocular depth estimation supported by elastic deformations can be made possible.

Reference symbol list

10

Schedule

12

Output image

14

Lidar data

16

Eyelid measurement points

18

transformed image

20

transformed lidar data

22

Shifted eyelid measurement points

S1-S7

Procedural steps

QUOTES INCLUDED IN THE DESCRIPTION

This list of documents listed by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Cited patent literature

• DE 102019103192 A1 [0003]
• DE 112020001897 C5 [0004]

Claims (10)

Method (10) for generating training data (18, 20) for training a machine learning model to estimate depth maps from mono camera images (12), wherein - a mono camera image (12) of a scene and depth data (14) corresponding to individual pixels between a respective object in the scene and a recording device are provided as ground truth data for the same scene, - At least one transformed camera image (18) with spatially shifted pixel values compared to the underlying mono camera image (12) is generated from the mono camera image (12) by a non-linear, location-dependent transformation, - transformed depth data (20) are generated from the corresponding depth data (14) using the same transformation, and - A transformed camera image (18) and the corresponding transformed depth data (20) serving as ground truth data for this transformed camera image (18) are provided as training data (18, 20). Procedure (10) according to Claim 1 , characterized in that eyelid measurement data (14) is used as depth data (14), which only indicates distances for a part of the pixels. Method (10) according to one of the preceding claims, characterized in that an elastic deformation is used as the transformation. Method (10) according to one of the preceding claims, characterized in that for the transformation - a noise image is generated with a number of pixels corresponding to the number of pixels of the mono camera image (12), the pixel values of which correspond to noise, in particular random or pseudo-random, - displacement data are generated from the noise image, in particular by means of a predetermined filter, which indicate a spatial displacement for each pixel, and - the pixel values of the mono camera image (12) are shifted in accordance with the displacements specified for this in the displacement data. Procedure (10) according to Claim 4 , characterized in that a predetermined Gaussian filter is applied to the noise image to generate the displacement data. Method (10) according to one of the preceding claims, characterized in that the shift of the pixel values is limited to a predetermined maximum value. Method (10) according to one of the preceding claims, characterized in that a plurality of correspondingly different training data (18, 20) is generated from a specific monocamera image (12) by means of several different transformations. Method (10) according to one of the preceding claims, characterized in that the and/or further training data (18, 20) are generated from the mono camera image (12) and the corresponding depth data (14) by applying at least one predetermined global linear transformation. Method (10) for training a machine learning model to estimate depth maps for mono camera images (12), where - at least one mono camera image (12) and depth data (14) corresponding to this are provided as ground truth data, - Training data (18, 20) are generated therefrom using a method (10) according to one of the preceding claims, and - The machine learning model is trained using the training data (18, 20) generated in this way, in particular using supervised learning. Assistance system for a motor vehicle, which has an interface for acquiring monocamera images (12) and a data processing device with a machine learning model according to a method (10). Claim 9 is trained, has and is set up to generate depth maps for captured mono camera images (12).

Images