Image compression and computer vision

Abstract

According to a computer-implemented method for image compression and computer vision, a camera image (14) is received and at least one feature map is generated by applying at least one trained residual neural network (7a) to the camera image (14). A compressed image (15) of the camera image (14) is generated by applying a trained first artificial neural network, ANN, (5) to first input data dependent on the at least one feature map, and a computer vision task is performed by applying a trained second ANN (6a, 6b, 6c, 6d) to second input data dependent on the at least one feature map.

Description

The present invention is directed to a computer-implemented method for image compression and computer vision and to a corresponding method for image decompression. The invention is further directed to a corresponding computer-implemented training method for training artificial neural networks. The invention is also directed to a corresponding method for at least partially automatically guiding a motor vehicle, to a data processing device, to an electronic vehicle guidance system which has such a data processing device, and to corresponding computer program products.Automotive cameras, for example, cameras with wide angle fisheye lenses, are commonly used to capture images for storage and transmission. It may be advantageous to compress these images in order to save storage space and transmission bandwidth with high image quality. Traditional compression techniques can cause significant distortion, particularly in the corners of fish eye images. Furthermore, low bit rate compression using these techniques may result in loss of essential information, which represents a potential safety risk, particularly for automotive applications.In the publication Z. Cheng et al.: "Learned Image Compression with Discrete Gaussian Mixture Likelihoods and Attenuation Modules" (arXiv:2001.01568v3), the authors describe how discretized Gaussian mixture probabilities can be used to parameterize the distributions of latent encodings for compression of images.The publication D. Minen et al.: "Joint Autoregressive and Hierarchical Priors for Learned Image Compression" (arXiv:1809.02736v 1) deals with autoregressive, hierarchical and combined priors as alternatives in connection with image compression.Furthermore, the uncompressed camera images can also be used to use one or more computer vision algorithms, for example to assist a driver or to automatically guide the motor vehicle.The publication H. Kaming et al.: "Deep Residual Learning for Image Recognition" (arXiv:1512.03385v 1) describes the concept of residual neural networks, also referred to as ResNet, which have convolutional layers with skip connections and can be trained for deep feature extraction of camera images.An object of the present invention is to provide a way in which image compression can be achieved and one or more computer vision tasks can be performed.This object is achieved by the subject matter of the independent claim. Further embodiments and preferred embodiments are subject matter of the dependent claims.The invention is based on the idea of extracting features from a camera image by using at least one residual neural network and using the extracted features both for image compression and for executing a computer vision task.According to one aspect of the invention, a computer-implemented method for image compression and computer vision is provided. In this case, a camera image, in particular a camera image which images an external environment of the motor vehicle, is received and at least one feature map is generated by applying at least one trained residual neural network to the camera image. A compressed image is generated by applying a trained first artificial neural network, ANN, to first input data, wherein the first input data depends on the at least one feature map. A computer vision task is performed by applying a trained second artificial neural network to second input data that depends on the at least one feature map.Unless otherwise stated, all steps of the computer-implemented method can be carried out by a data processing device which has at least one arithmetic logic unit, in particular a data processing device of the motor vehicle. In particular, the at least one computing unit is configured or adapted to carry out the steps of the computer-implemented method. For this purpose, the at least one computing unit can store, for example, a computer program which contains instructions which, when executed by the at least one computing unit, cause the at least one computing unit to carry out the computer-implemented method.All computing units of the at least one computing unit can be included in the motor vehicle. However, it is also possible for all computing units of the at least one computing unit to be part of an external computing system outside the motor vehicle, for example a backend server or a cloud computing system. It is also possible for the at least one computing unit to comprise at least one vehicle computing unit of the motor vehicle and at least one external computing unit which is contained by the external computing system. The at least one vehicle computing unit may be included, for example, by one or more electronic control units, ECUs (electronic control units), and / or one or more zone control units, ZCUs (zone control units), and / or one or more domain control units, DCUs (domain control units) of the motor vehicle and / or of the camera.For each execution of the computer-implemented method according to the invention, a respective execution of a method for image compression and computer vision, which is not purely computer-implemented, is obtained directly by including a step of generating the camera image, in particular by means of a camera, for example of the motor vehicle, in the method.The first input data or the second input data depends on the at least one feature map. In other words, the first input data is identical to the at least one feature map or is generated depending on the at least one feature map. Analogously, the second input data is identical to the at least one feature map or is generated as a function of the at least one feature map.The first ANN is in particular an ANN which is trained for image compression. The second ANN is, in particular, an ANN that is trained to execute the computer vision task. The architectures of the first and second ANN may be provided by or based on known architectures for neural networks for image compression and computer vision, respectively. The training of the first and the second ANN can also be based on known training approaches.In particular, the compressed image can be stored on a storage device and / or transmitted to a receiving unit, which can use or process the compressed image further. In particular, the compressed image can be used for driver assistance or for executing a computer vision task. Based on a result of the computer vision task and / or based on the compressed image, the motor vehicle can be guided at least partially automatically.Since, according to the invention, the at least one feature map is generated by means of at least one trained residual neural network, the extracted features of the at least one feature map are particularly suitable for executing both the computer vision task and the image compression. This is a significant advantage compared to alternative approaches where the extracted features are only well suited for image compression, but less well for computer vision tasks, or vice versa.According to some implementations of the computer-implemented method, one or more other computer vision tasks are performed. To execute one of the further computer vision tasks, a respective further trained second ANN is applied to the second input data or to the further second input data, which depends on the at least one feature map.According to some implementations, the at least one residual neural network includes at least one ResNet18 block. In particular, the ResNet18 block has 18 convolutional layers.According to at least one embodiment, at least one further feature map is generated by applying at least one further trained residual neural network to the at least one feature map, and the computer vision task is executed by applying the second ANN to the at least one further feature map.In other words, the second input data in this case consists of the at least one further feature map. The performance of the computer vision task can be further improved by the deeper feature extraction of such implementations.According to at least one embodiment, the compressed image is generated by applying the first ANN to the at least one feature map.In other words, the first input data in this case consists of the at least one feature map. By applying the at least one feature map instead of, for example, the at least one further feature map, loss of information due to image compression can be avoided or reduced.According to at least one embodiment, generating the compressed image includes generating at least one latent variable by applying an attention-based generative network of the first ANN to the at least one feature map.In other words, the compressed image is generated depending on the at least one latent variable.According to at least one embodiment, generating the compressed image includes generating at least one quantized variable by applying a quantization module of the first ANN to the at least one latent variable.In other words, the compressed image is generated depending on the at least one quantized variable.According to at least one embodiment, generating the compressed image includes generating at least one bitstream by applying an arithmetic encoder module of the first ANN to the at least one quantized variable.According to at least one embodiment, generating the compressed image includes generating at least one further quantized variable by applying a further quantization module of the first ANN to the at least one feature map.In other words, the compressed image is generated depending on the at least one further quantized variable.According to at least one embodiment, generating the compressed image includes generating at least one further bitstream by applying a further arithmetic encoder module of the first ANN to the at least one further quantized variable.According to at least one embodiment, at least one decoded variable is generated by applying an arithmetic decoder module of the first ANN to the at least one bitstream and applying a conditional prior module (ANN) of the first ANN to an output of the arithmetic decoder module.According to at least one embodiment, the at least one decoded variable is passed through an entropy parameter network of the first ANN.According to at least one embodiment, the at least one further decoded variable is generated by applying a further arithmetic decoder module of the first ANN to the at least one further bit stream depending on an output of the entropy parameter network.According to some embodiments, the reconstructed image is generated by applying a scale function module of the first ANN to the at least one further decoded variable.According to at least one embodiment of the computer-implemented method for image compression and computer vision, the second ANN is trained for object recognition and / or depth estimation and / or semantic segmentation and / or panoptic segmentation and / or motion estimation.In other words, the computer vision task includes an object detection task and / or a depth estimation task and / or a semantic segmentation task and / or a panoptic segmentation and / or motion estimation task.According to a further aspect of the invention, a method for at least partially automatically guiding a vehicle is provided. In this case, a computer-implemented method according to the invention for image compression and computer vision is carried out, in particular by the at least one arithmetic unit, for example of the motor vehicle.According to some embodiments, at least one control signal for at least partially automatically guiding the motor vehicle is generated, in particular by the at least one computing unit, depending on a result of the computer vision task and / or depending on the compressed image.The at least one control signal may be provided, for example, to one or more actuators of the motor vehicle, including, for example, one or more brake actuators and / or one or more steering actuators and / or one or more drive motors of the motor vehicle. The one or more actuators may influence a longitudinal and / or lateral control of the motor vehicle in order to at least partially automatically guide the motor vehicle.According to some embodiments, the assistance information for assisting the driver of the motor vehicle is generated, in particular by the at least one computing unit, depending on a result of the computer vision task and / or depending on the compressed image.The assistance information can be output by means of an output device of the motor vehicle, for example a display and / or an audio output system and / or a haptic output system.According to another aspect of the invention, there is provided a computer-implemented training method for training artificial neural networks to perform a computer-implemented method for image compression and computer vision according to the invention. In this case, the first ANN and the second ANN are trained using a common loss function. The common loss function is a weighted sum of a first loss function for training image compression, a second function for training the computer vision task, and an uncertainty loss function. In particular, respective weighting factors for the weighted sum can be trainable or trainable.According to some implementations of the computer-implemented method for image compression and computer vision according to the invention, the first ANN and the second ANN are trained by using a computer-implemented training method according to the invention.According to another aspect of the invention, a data processing apparatus is provided. The data processing device has at least one arithmetic unit which is configured to carry out a computer-implemented method according to the invention for image compression and computer vision and / or a computer-implemented training method according to the invention.A computing unit can be understood in particular as a data processing device which has processing circuits. The computing unit can thus process data in particular in order to carry out computing operations. This may also include operations to perform indexed accesses to a data structure, such as a look-up table, LUT.The computing unit can in particular contain one or more computers, one or more microcontrollers and / or one or more integrated circuits, for example one or more application-specific integrated circuits, ASICs (application-specific integrated circuit), one or more field programmable gate arrays, FPGAs, and / or one or more single-chip systems, SoCs (system on a chip). The computing unit may also include one or more processors, for example one or more microprocessors, one or more central processing units, CPUs (CPUs), one or more graphics processing units, GPUs (GPUs) and / or one or more signal processors, in particular one or more digital signal processors, DSPs. The computing unit may also include a physical or virtual group of computers or other of the aforementioned units.In various embodiments, the computing unit includes one or more hardware and / or software interfaces and / or one or more memory units.A memory unit can be used as volatile data memory, for example as dynamic random access memory, DRAM (dynamic random access memory) or static random access memory, SRAM (static random access memory), or as nonvolatile data memory, for example as read-only memory, ROM (read-only memory), as programmable read-only memory, PROM (programmable read-only memory), as erasable read-only memory, EPROM (erasable read-only memory), as electrically erasable read-only memory, EEPROM (electrically erasable read-only memory), as flash memory or flash EEPROM, as ferroelectric random access memory, FRAM (ferroelectric random access memory), as magnetoresistive random access memory, MRAM (magnetoresistive random access memory) or as phase-change random access memory, PCRAM (phase-change random access memory).According to another aspect of the invention, an electronic vehicle guidance system for a motor vehicle is provided. The electronic vehicle guidance system has a data processing device according to the invention, wherein the at least one computing unit is configured to carry out a method according to the invention for at least partially automatically guiding a motor vehicle.According to at least one embodiment, the electronic vehicle guidance system has the camera for the motor vehicle, which is configured to generate the camera image.According to another aspect of the invention, there is provided a computer program including instructions. If the instructions are executed by a data processing device, in particular by a data processing device according to the invention, the instructions cause the data processing device to carry out a computer-implemented method according to the invention for image compression and computer vision and / or a computer-implemented training method according to the invention.The instructions may be provided as program code, for example. The program code can be provided, for example, as binary code or assembler and / or as source code of a programming language, for example C, and / or as a program script, for example python.According to a further aspect of the invention, a further computer program comprising further instructions is provided. If the further commands are executed by an electronic vehicle guidance system according to the invention, in particular by a data processing device of the electronic vehicle guidance system, the further commands cause the electronic vehicle guidance system to carry out a method according to the invention for at least partially automatic guidance of a motor vehicle.The further instructions can be provided, for example, as program code. The program code can be provided, for example, as binary code or assembler and / or as source code of a programming language, for example C, and / or as a program script, for example python.According to a further aspect of the invention, a computer-readable storage medium is provided which stores a computer program according to the invention and / or a further computer program according to the invention.The computer program, the further computer program and the computer-readable storage medium can be referred to as respective computer program products that contain the instructions and / or the further instructions.Further features of the invention are evident from the claims, the figures and the description of the figures. The features and combinations of features mentioned above in the description and the features and combinations of features mentioned below in the description of the figures and / or shown in the figures can be encompassed by the invention not only in the respectively specified combination but also in other combinations. In particular, embodiments and combinations of features may also be included in the invention that do not have all of the features of an originally formulated claim. Moreover, embodiments and combinations of features may be encompassed by the invention that go beyond or depart from the combinations of features set forth in the appended claims.The invention is explained in detail below with reference to specific exemplary embodiments and respective schematic drawings. In the drawings, identical or functionally identical elements can be denoted by the same reference numerals. The description of identical or functionally identical elements is not necessarily repeated with reference to different figures.The figures show: FIG. 1 schematically shows a motor vehicle having an exemplary embodiment of an electronic motor vehicle guidance system according to the invention; FIG. 2 shows a schematic block diagram of an exemplary embodiment of a computer-implemented method for image compression and computer vision according to the invention; FIG. 3 shows a schematic block diagram of a further exemplary embodiment of a computer-implemented method according to the invention for image compression and computer vision; FIG. 4 shows a schematic block diagram of a further exemplary embodiment of a computer-implemented method according to the invention for image compression and computer vision; and FIG. 5 shows a schematic block diagram of a further exemplary embodiment of a computer-implemented method according to the invention for image compression and computer vision.FIG. 1 schematically shows a motor vehicle 1 with an exemplary embodiment of an electronic vehicle guidance system 3 according to the invention. The electronic vehicle guidance system 3 has a data processing device 4 with at least one computing unit and, for example, a camera 2 for the motor vehicle.The at least one computing unit is configured to execute a computer-implemented method according to the invention for image compression and computer vision. In particular, the at least one computing unit is configured to carry out an inventive method for at least partially automatically guiding a motor vehicle, which method includes carrying out an inventive computer-implemented method for image compression and computer vision.A schematic block diagram of an exemplary embodiment of a computer-implemented method for image compression and computer vision according to the invention is shown in FIG. 2.The at least one computing unit receives a camera image 14 generated by the camera 2 and generates at least one feature map by applying at least one trained residual neural network 7 ato the camera image 14.Furthermore, the at least one computing unit executes one or more computer vision tasks depending on the at least one feature map. For each computer vision task, the at least one computing unit applies a respective trained second ANN 6 a, 6 b, 6 c, 6 dto second input data that depend on the at least one feature map. In particular, the second input data can be given by the at least one feature map or by at least one further feature map, which is generated by applying at least one further trained residual neural network 7 bto the at least one feature map.Each of the second ANNs 6 a, 6 b, 6 c, 6 dmay be trained to perform a different computer vision task including, but not limited to, object detection, depth estimation, semantic segmentation, panoramic segmentation, and / or motion estimation, for example.The at least one computing unit can generate at least one control signal for at least partially automatically guiding the motor vehicle 1 and / or depending on a result of the computer vision task and / or depending on the compressed image 15.In FIG. 3, the first ANN 5 is shown in detail for exemplary embodiments of a computer-implemented method for image compression and computer vision according to the invention.In this case, at least one latent variable is generated by applying an attention-based generative network 8 of the first ANN 5 to the at least one feature map, and at least one quantized variable is generated by applying a quantization module 9 of the first ANN 5 to the at least one latent variable. A bitstream 21 is generated by applying an arithmetic coder module 16a of the first ANN 5 to the at least one quantized variable.At least one further quantized variable is generated by applying a further quantization module 12 of the first ANN 5 to the at least one feature map. A further bit stream 22 is generated by applying a further arithmetic coder module 18 aof the first ANN 5 to the at least one further quantized variable.To generate the compressed image 15, at least one decoded variable is generated by applying an arithmetic decoder module 16 bof the first ANN 5 to the bitstream 21. The arithmetic encoder module 16 aand the arithmetic decoder module 16 brepresent, for example, an entropy bottleneck module 16. A conditional prior module 10 of the first ANN 5 is applied to an output of the arithmetic decoder module 16b. The at least one decoded variable is passed through an entropy parameter network 17 of the first ANN 5. At least one further decoded variable is generated by applying a further arithmetic decoder module 18 bof the first ANN 5 to the further bit stream 22 depending on an output of the entropy parameter network 17. The further arithmetic encoder module 18 aand the further arithmetic decoder module 18 brepresent, for example, a generative channel module 18. A context model module 11 may be coupled between the further quantization module 12 and the conditional prior module 10. The reconstructed image 15 is generated by applying a scale function module 13 of the first ANN 5 to at least one decoded variable.FIG. 4 shows the first ANN 5 in detail for further exemplary embodiments. For example, the at least one residual neural network 7 aincludes two ResNet 18 blocks 19 a, 19 band the at least one further residual neural network 7 bincludes three further ResNet 18 blocks 19 c, 19 d, 19 e. The attention-based generative network 8 and the conditional prior module 10 each have, for example, five convolutional layers, and the entropy parameter network 17 may have three convolutional layers. The scale function module 13 may comprise an attenuation module followed by a sequence of eight convolutional layers followed by another attenuation module followed by another sequence of seven convolutional layers. Skip connections may be provided, for example, as outlined.Further details concerning individual components of the first ANN 5 can be found in the publications by Minen et al. and Cheng et al. mentioned at the beginning of the present disclosure.FIG. 5 shows a schematic block diagram of a further exemplary embodiment of a computer-implemented method for image decompression according to the invention.In this case, the second ANNs 6 a, 6 b, 6 c, 6 dmay be configured, for example, as a depth estimation network 6 a, also referred to as a distance estimation network, which provides a depth map 20 aof the camera image 14 as an output, a semantic segmentation network 6 b, which provides a semantic map 20 bof the camera image 14 as an output, a motion estimation network 6 c, which provides a motion map 20 cof the camera image 14 as an output, and an object recognition network 6 d, which provides object recognition data 20 das an output.According to some implementations, an autoregressive model, which may be referred to as a JointGAD, common Gaussian attenuation decoder is provided that uses the attenuation-based generative network 8 to generate the latent variable for image reconstruction. The attention-based generative network 8 may advantageously utilize the ResNet 18 model to extract meaningful features from the input camera image 14, and the attention mechanism helps the model focus on relevant areas of the camera image 14. The generated latent variable is then quantized and passed through an entropy matching module 16 which compresses the data to a lower bit rate. The compressed data is then decoded back into the latent variable and used to reconstruct the camera image 14 using the generative channel module 18. The reconstructed image 15 has a low bit rate and high visual quality, particularly at the corners of the fish eye images, thereby being suitable for efficient storage and transmission in automotive applications.According to some implementations, an advanced image compression technique is provided that combines the performance of an attention-based generative network with a common Gaussian decoder to address the technical problem of efficiently compressing images while maintaining high image quality, particularly for fish-eye images covering a wide field of view, while running in conjunction with other computer vision tasks.Some embodiments of the invention may overcome or partially overcome various disadvantages of other compression approaches, including, for example, limited compression performance, inability to compress complex scenes, artifacts in the compressed image, significant loss of information, and / or significant computational requirements.In accordance with some embodiments of the invention, the attention-based generative network 8 is used to generate a latent variable that is quantized using the quantization module 9. The quantization module 9 maps the continuous latent variable into a discrete latent variable. This quantized variable is then encoded using a hyper-prior model and the resulting bitstreams 21 are passed through the entropy matching module 16. The hyper-prior model generates a sequence of scales and means that are used in the entropy matching module 16 to encode the quantized latent variable. The encoded quantized latent variable is passed through the entropy matching module 16 to produce a sequence of bitstreams 21. The bitstreams 21 are then decoded using a conditional prior model and the resulting decoded variable is passed through the generative channel module 18 to reconstruct the input camera image 14. The conditional prior model generates a sequence of scales and means used to decode the bitstreams back into the discrete latent space. The generative channel module 18 generates a sequence of scales and means used to reconstruct the camera image 14 by the scale function module 13. The resulting reconstructed image 15 may have a significantly higher quality than is achieved by alternative compression techniques.According to some implementations, a variation autoencoder approach may be used to train the ANNs 5, 6 a, 6 b, 6 c, 6 d, 7 a, 7 b. The encoder is, for example, a ResNet 18 model which compresses the input image into a low-sized latent space, while the decoder is a network architecture as described above. The computer vision tasks are based on typical decoder architectures that learn together with the compression decoder according to a multi-task learning approach. During training, the ANNs 5, 6a, 6b, 6c, 6d, 7a, 7b are optimized to minimize the difference between the input image and the reconstructed image by adding the loss of computer vision tasks. This results in an efficient and accurate reconstruction of high quality images.In other words, in addition to a dedicated loss function for compression, for example a mean square loss that is not passed through the quantized bit optimization after the entropy matching module 16, separate loss functions may be used for each computer vision task. These task specific losses guide the learning process to generate features effective for the respective computer vision tasks. For example, a self-monitoring Lovasz softmax loss may be used for a semantic segmentation task, and the Lovasz softmax and focus loss may be used for motion segmentation. A YOLOv3loss may be used for object detection, and an loUloss may be added using a segmentation mask. Furthermore, an overall loss, also referred to as uncertainty loss, may be used to further improve the overall performance of the architecture.A central feature of some implementations is the use of a common Gaussian decoder that is capable of capturing higher order statistical dependencies in the camera image 14. This allows the model to better preserve important information, particularly in fish eye images commonly used in automotive applications. Compared to traditional compression techniques, a substantially higher image quality can be achieved, especially for fish eye images.

Claims

A computer-implemented method for image compression and computer vision, wherein a camera image (14) is received, and wherein - at least one feature map is generated by applying at least one trained residual neural network (7a) to the camera image (14); - a compressed image (15) is generated by applying a trained first artificial neural network, ANN, (5) to first input data depending on the at least one feature map; and - a computer vision task is executed by applying a trained second ANN (6a, 6b, 6c, 6d) to second input data depending on the at least one feature map.The computer-implemented method of claim 1, wherein the at least one residual neural network (7a) comprises at least one ResNet18 block (19a, 19b).The computer-implemented method according to any one of the preceding claims, wherein - at least one further feature map is generated by applying at least one further trained residual neural network (7b) to the at least one feature map, and the computer vision task is executed by applying the second ANN (6a, 6b, 6c, 6d) to the at least one further feature map; and / or - the compressed image (15) is generated by applying the first ANN (5) to the at least one feature map.The computer-implemented method of any preceding claim, wherein generating the compressed image (15) includes generating at least one latent variable by applying an attention-based generative network (8) of the first ANN (5) to the at least one feature map.The computer-implemented method of claim 4, wherein generating the compressed image (15) includes generating at least one quantized variable by applying a quantization module (9) of the first ANN (5) to the at least one latent variable.The computer-implemented method of claim 5, wherein generating the compressed image (15) includes generating at least one bitstream (21) by applying an arithmetic encoder module (16a) of the first ANN (5) to the at least one quantized variable.The computer-implemented method according to any of the preceding claims, wherein generating the compressed image (15) includes generating at least one further quantized variable by applying a further quantization module (12) of the first ANN (5) to the at least one feature map.The computer-implemented method of claim 7, wherein generating the compressed image (15) includes generating at least one further bitstream (22) by applying a further arithmetic encoder module (18a) of the first ANN (5) to the at least one further quantized variable.The computer-implemented method according to claim 6 and claim 8, wherein - at least one decoded variable is generated by applying an arithmetic decoder module (16b) of the first ANN (5) to the at least one bit stream (21) and applying a conditional prior module (10) of the first ANN (5) to an output of the arithmetic decoder module (16b); - the at least one decoded variable is passed through an entropy parameter network (17) of the first ANN (5); - at least one further decoded variable is generated depending on an output of the entropy parameter network (17) by applying a further arithmetic decoder module (18b) of the first ANN (5) to the at least one further bit stream (22); and - the compressed image (15) is generated by applying a scale function module (13) of the first ANN (5) to the at least one further decoded variable.The computer-implemented method according to any one of the preceding claims, wherein the second ANN (6a, 6b, 6c, 6d) is trained for object recognition and / or depth estimation and / or semantic segmentation and / or panoptic segmentation and / or motion estimation.A computer-implemented training method for training artificial neural networks to perform a computer-implemented method according to any of the preceding claims, wherein the first ANN (5) and the second ANN (6a, 6b, 6c, 6d) are trained using a common loss function which is a weighted sum of a first loss function for training the computer vision task, a second loss function for training image compression and an uncertainty loss function.The computer-implemented method of any one of claims 1 to 10, wherein the first ANN (5) and the second ANN (6a, 6b, 6c, 6d) are trained using a computer-implemented training method of claim 11.Method for at least partially automatic guidance of a motor vehicle (1), wherein a computer-implemented method according to one of Claims 1 to 10 or 12 is executed and - at least one control signal for at least partially automatic guidance of the motor vehicle (1) is generated as a function of a result of the computer vision task and / or as a function of the compressed image (15); and / or - assistance information for assisting a driver of a motor vehicle (1) is generated as a function of a result of the computer vision task and / or as a function of the compressed image (15).Data processing device (4) comprising at least one computing unit which is configured to carry out a computer-implemented method according to one of Claims 1 to 10 or 12 and / or a computer-implemented training method according to Claim 11.Electronic vehicle guidance system (3) comprising a data processing device (4) according to claim 14, wherein the at least one computing unit is configured to carry out a method for at least partially automatically guiding a motor vehicle (1) according to claim 13.Electronic vehicle guidance system (3) according to Claim 15, having a camera (2) for the motor vehicle (1) which is configured to generate the camera image (14).A computer program product comprising - instructions which, when executed by a data processing device (4), cause the data processing device (4) to perform a computer-implemented method according to any one of claims 1 to 10 or 12 and / or a computer-implemented training method according to claim 11; and / or - further instructions which, when executed by an electronic vehicle guidance system (3) according to any one of claims 15 or 16, cause the electronic vehicle guidance system (3) to perform a method according to claim 13.

Images