Computer-implemented method for generating synthesized data based on an adversarial neural network

Abstract

The invention relates to a computer-implemented method for generating synthesized data (24) of a synthetic scene and / or environment with a plurality of three-dimensional synthetic objects and / or entities based on an adversarial neural network (10, 10a, 10b), the adversarial neural network (10, 10a, 10b) including a generator module (12) and discriminator module (14), the generator module (12) and the discriminator module (14) contesting with each other in form of a zero-sum game, where a loss function (16) of one of the generator module (12) and the discriminator module (14) corresponds to a gain function (16) of the other one of the generator module (12) and the discriminator module (14), the generator module (12) being configured for generating synthetic data (18) of the synthetic scene and / or environment based on a plurality of latent variables (20), each latent variable (20) corresponding to one of the plurality of synthetic objects and / or entities, the discriminator module (14) being configured for discriminating the synthetic data (18) with respect to reference data (22) based on a set of reference scenes and / or environments with a plurality of reference objects and / or entities.

Description

[0001] DESCRIPTION

[0002] Computer-implemented method for generating synthesized data based on an adversarial neural network

[0003] TECHNICAL FIELD

[0004] The invention relates to a computer-implemented method for generating synthesized data based on an adversarial neural network. The invention further relates to a computing device, a computer program, and a computer-readable storage medium.

[0005] BACKGROUND

[0006] Generative adversarial networks provide a framework for estimating generative models via an adversarial process, in which two models, a generative model G and a discriminative model D are trained simultaneously. Generative adversarial networks may be used to generate synthesized data of a plurality of synthetic scenes and / or environments. The synthetic scene and / or environment may show an image of a synthetic human face and / or a plurality of synthetic objects and / or entities. The synthesized data may be used to train other machine learning algorithms. Thus, the generation of training data by annotating, classifying, and / or labelling data by humans may be avoided or available training data may be complemented with the synthesized data.

[0007] SUMMARY OF THE INVENTION

[0008] The object of the invention is to provide an improved method for generating synthesized data of a plurality of synthetic scenes and / or environments.

[0009] The object of the invention is achieved by the subject-matter of the independent claims. Advantageous embodiments of the invention are subject-matter of the dependent claims. In one aspect, the invention provides a computer-implemented method for generating synthesized data of a synthetic scene and / or environment with a plurality of three-dimensional synthetic objects and / or entities based on an adversarial neural network, the adversarial neural network including a generator module and discriminator module, the generator module and the discriminator module contesting with each other in form of a zero-sum game, where a loss function of one of the generator module and the discriminator module corresponds to a gain function of the other one of the generator module and the discriminator module, the generator module being configured for generating synthetic data of the synthetic scene and / or environment based on a plurality of latent variables, each latent variable corresponding to one of the plurality of synthetic objects and / or entities, the discriminator module being configured for discriminating the synthetic data with respect to reference data based on a set of reference scenes and / or environments with a plurality of reference objects and / or entities, the method comprising: a) generating, by the generator module, three-dimensional synthetic depth data for each of the plurality of latent variables, the synthetic depth data being indicative of a synthetic depth of the corresponding synthetic object and / or entity; b) discriminating, by the discriminator module, the synthetic depth data with respect to three-dimensional reference depth data, the reference depth data being indicative of a respective reference depth for each of the reference objects and / or entities; c) training and / or updating the generator module and the discriminator module, respectively, with respect to a loss and / or gain function depending on the synthetic depth data and the reference depth data; and d) generating synthesized data of a plurality of synthetic scenes and / or environments by using the trained and / or updated generator module and / or discriminator module.

[0010] An advantage of the method may be that a three-dimensional depth is considered in the zero-sum game of the generator module and the discriminator module. Thus, the generator module and the discriminator module are trained and / or updated with respect to three dimensions. This improves the quality of the synthesized data and enables to generate three-dimensional training data based on the synthesized data.

[0011] Preferably, the adversarial neural network is based on a generative adversarial network, a generative adversarial transformer, an adversarial variational autoencoder, a vision-based transformer, a diffusion model, and / or a combination thereof.

[0012] Preferably, step c) comprises: c1) training and / or updating such that the synthetic depth data is indicative of and / or corresponds to the reference depth data.

[0013] An advantage of the method may be that the loss and / or gain function depends on the synthetic depth data and the reference depth data. The generator module and the discriminator module can thus be trained and / or updated such that the synthetic depth data is indicative of and / or corresponds to the reference depth data. This improves the quality of the synthesized data.

[0014] Preferably, step d) comprises: d1) generating three-dimensional synthesized ground-truth data; and / or d2) generating two-dimensional synthesized depth map data.

[0015] An advantage of the method may be that three-dimensional synthesized groundtruth data may be generated. The three-dimensional synthesized ground-truth data may be used for training three-dimensional machine learning algorithms such as three-dimensional object detection algorithms. This improves the training quality of such algorithms. Another advantage of the method may be that two-dimensional synthesized depth map data may be generated. The two-dimensional synthesized depth map data may be used for training two-dimensional machine learning algorithms such as depth estimation algorithms. This improves the training quality of such algorithms.

[0016] Preferably, the method further comprises: e1) based on the synthesized data, training a machine learning algorithm for analyzing sensor data provided by a sensor device for perceiving a scene and / or environment thereof.

[0017] The synthesized data may be used for training a machine learning algorithm for analyzing sensor data provided by a sensor device for perceiving a scene and / or environment. The sensor device may be, for example, a camera, lidar, and / or radar device. Since the quality of the synthesized data is improved, the quality of the training of the machine learning algorithm based on the synthesized data and / or the results of the analysis of the sensor data are improved as well.

[0018] Preferably, the sensor device is included in an autonomous vehicle for perceiving a scene and / or environment thereof and the method further comprises: e2) causing the autonomous vehicle to navigate based on the analysis.

[0019] The trained machine learning algorithm may be configured for analyzing sensor data provided by a sensor device in an autonomous vehicle. In this case, the method may include causing the autonomous vehicle to navigate based on the analysis. The autonomous vehicle may any kind of vehicle, such as a car, robot, aircraft, watercraft, etc. The causing may include generating a navigation signal for navigating the autonomous vehicle.

[0020] Preferably, the method comprises one, several, or all of the following: f1) based on the synthesized data, training a three-dimensional object-detection algorithm for detecting one or more objects and / or entities from sensor data provided by a sensor device for perceiving a scene and / or environment thereof; and / or f2) based on the synthesized data, training a depth estimation algorithm for estimating a depth of an object and / or entity from sensor data provided by a sensor device for perceiving a scene and / or environment thereof; and / or f3) based on the synthesized data, training an instance segmentation algorithm for distinguishing between instances of objects and / or entities of same category; and / or f4) based on the synthesized data, training a semantic segmentation algorithm for distinguishing between objects and / or entities of different categories; and / or f5) based on the synthesized data, training a panoptic segmentation algorithm for distinguishing between instances of objects and / or entities of same and / or different categories.

[0021] The machine learning algorithm may be, for example, a three-dimensional objectdetection algorithm, a depth estimation algorithm, an instance segmentation algorithm, a semantic segmentation algorithm, and / or a panoptic segmentation algorithm for analyzing sensor data provided by a sensor device for perceiving a scene and / or environment thereof.

[0022] Preferably, the method comprises one, several, or all of the following: g1) adapting the synthesized data to one or more parameters of a sensor device for perceiving a scene and / or environment thereof; and / or g2) projecting the synthesized data from three dimensions on two dimensions or vice versa; and / or g3) based on the synthesized data, creating three-dimensional point cloud-based data, ; and / or g4) based on the synthesized data, creating two-dimensional image data.

[0023] An advantage of the method may be that the synthesized data may be adapted to one or more parameters of the sensor device. The sensor data may depend on one or more sensor parameters, such as field of view, range, resolution, and so on of the sensor device. The method may generate synthesized data that is adapted to the one or more sensor parameters. Thus, the three-dimensional objectdetection algorithm can be trained and / or adapted to the sensor device.

[0024] Another advantage may be that the method may generate three-dimensional point cloud-based data. Thus, the synthesized data may also be used to train machine learning algorithms that process point cloud-based data that may be provided by the sensor device.

[0025] Preferably, the synthetic depth data and / or reference depth data include: hi) one or more three-dimensional point variables for each of the synthetic objects and / or entities and / or of the reference objects and / or entities, respectively; and / or h2) a three-dimensional bounding box variable for each of the synthetic objects and / or entities and / or of the reference objects and / or entities, respectively.

[0026] The depth in the synthetic depth data and / or the reference depth data may, for example, be provided by three-dimensional point variables for each of the objects and / or entities. For example, the depth may be provided by a three-dimensional bounding box variable. The bounding box variable may include eight points of a bounding box around the corresponding object and / or entity. Thus, it may be sufficient to configure the adversarial neural network to additionally process a limited number of point and / or bounding box variable. Thus, the method may still perform efficiently.

[0027] Preferably, the generating and / or discriminating includes predicting the one or more three-dimensional point variables and / or the three-dimensional bounding box variable of the synthetic depth data and / or reference depth data, preferably by using a three-dimensional object detector algorithm.

[0028] Preferably, the loss and / or gain function is configured:

[0029] 11) as a L1 and / or L2 loss and / or gain function; and / or

[0030] 12) for monocular depth estimation.

[0031] In another aspect, the invention provides a computing device comprising means adapted to execute the steps of the method according to any of the preceding embodiments.

[0032] In another aspect, the invention provides a computer program comprising instructions to cause a computing device to execute the steps of the method according to any of the preceding embodiments.

[0033] In another aspect, the invention provides a computer-readable storage medium having stored thereon the computer program. Embodiments of the invention are based on an adversarial neural network as described in the following documents:

[0034] [1] Ian J. Goodfellow et aL: “Generative Adversarial Networks”; 2014; hit ns: / / arxiv.org / abs / 1406.2661 ;

[0035] [2] Tero Karras et aL: “A Style-Based Generator Architecture for Generative

[0036] Adversarial Networks”; 2019; https: / / arxiv.org / abs / 1812.04948;

[0037] [3] Drew A. Hudson et aL: “Generative Adversarial Transformers”; 2022; https: / / arxiv.org / abs / 2103,01209; and

[0038] [4] Drew A. Hudson and C. Lawrence Zitnick: “Compositional Transformers for

[0039] Scene Generation”; 2021 ; https: / / arxiv.org / abs / 2111 ,08960.

[0040] Preferred embodiments of the invention have the following advantages and effects:

[0041] - 3D ground-truths are available for objects generated in the synthetic images, allowing it to be used for camera-based 3D object detection; and / or

[0042] - the projection between 3D-2D conform to the specific camera parameters of the data provided; and / or

[0043] - the generated depth maps can be used as they are regressed to be within a real parameter space instead.

[0044] BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Embodiments of the invention are now explained in more detail with reference to the accompanying drawings of which

[0046] Fig. 1 shows a comparative embodiment of an adversarial neural network;

[0047] Fig. 2 shows an embodiment of the adversarial neural network according to the invention; and Fig. 3 shows an embodiment of a computer-implemented method for generating synthesized data of a synthetic scene and / or environment based on the adversarial neural network according to Fig. 2.

[0048] DETAILED DESCRIPTION OF THE DRAWINGS

[0049] Fig. 1 shows a comparative embodiment of an adversarial neural network 10.

[0050] The adversarial neural network 10 includes a generator module 12 and a discriminator module 14. In the adversarial neural network 10, the generator module 12 and the discriminator module 14 contest with each other in form of a zero-sum game, where a loss function 16 of one of the generator module 12 and the discriminator module 14 corresponds to a gain function 16 of the other one of the generator module 12 and the discriminator module 14.

[0051] The generator module 12 is configured for generating synthetic data 18 of a synthetic scene and / or environment with a plurality of three-dimensional synthetic objects and / or entities. The generator module 12 generates the synthetic data 18 based on a plurality of latent variables 20, each latent variable 20 corresponding to one of the plurality of synthetic objects and / or entities.

[0052] The discriminator module 14 is configured for discriminating the synthetic data 18 with respect to reference data 22 based on a set of reference scenes and / or environments with a plurality of reference objects and / or entities.

[0053] In the adversarial neural network 10, the generator module 12 and the discriminator module 14 are trained and / or updated with respect to a loss and / or gain function 16 depending on the synthetic data 18 and the reference data 22.

[0054] The generator module 12 is trained and / or updated to generate or produce the synthetic data 18 with respect to the loss and / or gain function 16 with increasing similarity to the reference data 22, while the discriminator module 14 is trained and / or updated to discriminate or challenge the synthetic data 18 with respect to the loss and / or gain function 16 with increasing requirements. The adversarial neural network 10 may be trained and / or updated recurrently. Once trained and / or updated, the adversarial neural network 10 may be used to generate synthesized or segment data 24 of a plurality of synthetic scenes and / or environments.

[0055] Fig. 2 shows an embodiment of the adversarial neural network 10 according to the invention.

[0056] The adversarial neural network 10 includes the features of the comparative embodiment according to Fig. 1 .

[0057] In contrast to the comparative embodiment, however, the generator module 12 is configured for generating three-dimensional synthetic depth data 26 for each of the plurality of latent variables 20, the synthetic depth data 26 being indicative of a synthetic depth of the corresponding synthetic object and / or entity.

[0058] Furthermore, the discriminator module 14 is configured for discriminating the synthetic depth data 26 with respect to three-dimensional reference depth data 28, the reference depth data 28 being indicative of a respective reference depth for each of the reference objects and / or entities.

[0059] The synthetic depth data 26 and / or the reference depth data 28 may include one or more three-dimensional point variables for each of the synthetic objects and / or entities and / or of the reference objects and / or entities, respectively. The synthetic depth data 26 and / or the reference depth data 28 may, for example, a three- dimensional bounding box variable for each of the synthetic objects and / or entities and / or of the reference objects and / or entities.

[0060] The generating and / or the discriminating may include a prediction of the one or more three-dimensional point variables and / or of the three-dimensional bounding box variable of the synthetic depth data and / or the reference depth data, preferably by using a three-dimensional object detector algorithm. In the adversarial neural network 10, the generator module 12 and the discriminator module 14 are trained and / or updated with respect to the loss and / or gain function 16 - in addition or in alternative to the comparative embodiment - depending on the synthetic depth data 26 and the reference depth data 28.

[0061] The adversarial neural network 10 may be trained and / or updated such that the synthetic depth data 28 is indicative of and / or corresponds to the reference depth data 28. The loss and / or gain function 16 may be configured as a L1 and / or L2 loss and / or gain function. Additionally, or alternatively, the loss and / or gain function 16 may be configured for monocular depth estimation. Furthermore, the loss and / or gain function 16 may come from the point variables for the reference objects and / or entities obtained by the three-dimensional object detection algorithm against the point variables for the synthetic objects and / or entities generated by the generator module 12.

[0062] Fig. 3 shows an embodiment of a computer-implemented method for generating the synthesized data 24 of the synthetic scene and / or environment based on the adversarial neural network 10 according to Fig. 2.

[0063] In the method as shown in Fig. 2, the latent variables 20 are fed into a machine learning algorithm in order to produce structure latents 30 indicative of a shape, depth and / or position of each of the synthetic objects and / or entities in the synthetic scene.

[0064] The structure latents 30 are then fed into a first adversarial neural network 10a according to Fig. 2. The first adversarial neural network 10a generates the synthesized data 24. The synthesized data 24 may, for example, include three- dimensional synthesized ground-truth data 32, two-dimensional depth map data 34, and / or two-dimensional semantic map data 36.

[0065] The depth map data 34 may be projected from two dimensions on three dimensions in order to create three-dimensional point cloud-based data 38. The structure latents 30, the depth map data 34 and / or the semantic map data 36 may also be projected from three dimensions on two dimensions in order to create two-dimensional image data 40. For this, the first adversarial neural network 10a is trained and / or updated recurrently and the structure latents 30 are fed into another machine learning algorithm in order to produce style latents 42 indicative of a style of each of the synthetic objects and / or entities of the synthetic scene and / or environment. The style latents 42, the depth map data 34, and the semantic map data 36 are then fed into a second adversarial neural network 10b according to Fig. 2. The second adversarial neural network 10b thus generates the image data 38.

[0066] The synthesized data 24 may be used to train a three-dimensional objectdetection algorithm for detecting one or more objects and / or entities from sensor data provided by a sensor device. The synthesized data 24 may further be used to train a depth estimation algorithm for estimating a depth of an object and / or entity from the sensor data provided by the sensor device. Therefore, the synthesized data 24 may be adapted to one or more parameters of the sensor device.

[0067] The synthesized data 24 may be used to train an instance segmentation algorithm. Machine learning algorithms for instance segmentation enable to distinguish between instances of objects and / or entities of same category. Since the synthesized data 24 accounts for three-dimensional synthetic and reference depth data 26, 28, the synthesized data 24 may be used to train the instance segmentation algorithm in order to distinguish between instances of objects and / or entities of same category having different synthetic and / or reference depth.

[0068] Additionally, or alternatively, the synthesized data 24 may be used to train a semantic segmentation algorithm. Machine learning algorithms for semantic segmentation enable to distinguish between objects and / or entities of different categories. Since the synthesized data 24 accounts for three-dimensional synthetic and reference depth data 26, 28, for example in form of the semantic map data 36, the synthesized data 24 may be used to train the semantic segmentation algorithm in order to distinguish between objects and / or entities of different categories having different synthetic and / or reference depth. Additionally, or alternatively, the synthesized data 24 may be used to train a panoptic segmentation algorithm. Machine learning algorithms for panoptic segmentation can solve both, instance segmentation and semantic segmentation. Since the synthesized data 24 accounts for three-dimensional synthetic and reference depth data 26, 28, the synthesized data 24 may be used to train the panoptic segmentation algorithm in order to distinguish between instances of objects and / or entities of same and / or different categories having different synthetic and / or reference depth.

[0069] Preferred embodiments may be summarized as follows:

[0070] Camera-based 3D object detection architectures tend to be tied to the specific camera parameters for 3D-2D projections. A further problem may be insufficient annotated data for 3D camera object detection that uses specific camera parameters. Furthermore, synthetic image generation architectures do not generate required 3D data for specific instances of objects in the image.

[0071] Preferred embodiments propose to generate synthetic images and 3D groundtruths through use of at least one generative adversarial transformer, such as described in document [4], but in a modified version:

[0072] Preferred modifications of the generator are:

[0073] - addition of generated 3D data per each segment created from each structure latent, which could be the 8 points in a 3D bounding box; and / or

[0074] - supervise the depth map generation using additional losses input into the discriminator; and / or

[0075] - projection of the generated depth map to create a pseudo-lidar point cloud using the specific camera parameters of the dataset.

[0076] Preferred modifications of the discriminator are:

[0077] - usage of a monocular depth estimator to supervise the depth map generated, providing a loss, such as a point-wise L1 loss, to the discriminator; and / or - usage of a point 3D object detector to predict the 3D bounding boxes of both the real samples from the dataset and fake ones generated, providing a Euclidean loss based on the 3D bounding boxes to the discriminator; and / or

[0078] - the 3D bounding box loss is masked over segment instances of relevant objects.

[0079] The invention provides the computer-implemented method for generating the synthesized data 24 of the synthetic scene and / or environment with the plurality of three-dimensional synthetic objects and / or entities based on the adversarial neural network 10. The invention further provides a computing device (not shown) comprising means adapted to execute the steps of the described method. The invention further provides a computer program (not shown) comprising instructions to cause the computing device to execute the steps of the described method. The invention further provides a computer-readable storage medium (not shown) having stored thereon the computer program.

[0080] REFERENCE SIGNS

[0081] 10 adversarial neural network

[0082] 10a first adversarial neural network

[0083] 10b second adversarial neural network

[0084] 12 generator module

[0085] 14 discriminator module

[0086] 16 loss and / or gain function

[0087] 18 synthetic data

[0088] 20 latent variable

[0089] 22 reference data

[0090] 24 synthesized data

[0091] 26 synthetic depth data

[0092] 28 reference depth data

[0093] 30 structure latent

[0094] 32 ground-truth data

[0095] 34 depth map data

[0096] 36 semantic map data

[0097] 38 point cloud-based data

[0098] 40 image data

[0099] 42 style latent

Claims

CLAIMS1 . A computer-implemented method for generating synthesized data (24) of a synthetic scene and / or environment with a plurality of three-dimensional synthetic objects and / or entities based on an adversarial neural network (10, 10a, 10b), the adversarial neural network (10, 10a, 10b) including a generator module (12) and discriminator module (14), the generator module (12) and the discriminator module (14) contesting with each other in form of a zero-sum game, where a loss function (16) of one of the generator module (12) and the discriminator module (14) corresponds to a gain function (16) of the other one of the generator module (12) and the discriminator module (14), the generator module (12) being configured for generating synthetic data (18) of the synthetic scene and / or environment based on a plurality of latent variables (20), each latent variable (20) corresponding to one of the plurality of synthetic objects and / or entities, the discriminator module (14) being configured for discriminating the synthetic data (18) with respect to reference data (22) based on a set of reference scenes and / or environments with a plurality of reference objects and / or entities, the method comprising: a) generating, by the generator module (12), three-dimensional synthetic depth data (26) for each of the plurality of latent variables (20), the synthetic depth data (26) being indicative of a synthetic depth of the corresponding synthetic object and / or entity; b) discriminating, by the discriminator module (14), the synthetic depth data (26) with respect to three-dimensional reference depth data (28), the reference depth data (28) being indicative of a respective reference depth for each of the reference objects and / or entities; c) training and / or updating the generator module (12) and the discriminator module (10), respectively, with respect to a loss and / or gain function (16) depending on the synthetic depth data (26) and the reference depth data (28); and d) generating synthesized data (24) of a plurality of synthetic scenes and / or environments by using the trained and / or updated generator module (12) and / or discriminator module (14).

2. The method according to claim 1 , characterized in that the adversarial neural network (10, 10a, 10b) is based on a generative adversarial network, a generativeadversarial transformer, an adversarial variational autoencoder, a vision-based transformer, a diffusion model, and / or a combination thereof.

3. The method according to any of the preceding claims, characterized in that step c) comprises: c1) training and / or updating such that the synthetic depth data (26) is indicative of and / or corresponds to the reference depth data (28).

4. The method according to any of the preceding claims, characterized in that step d) comprises: d1) generating three-dimensional synthesized ground-truth data (32); and / or d2) generating two-dimensional synthesized depth map data (34).

5. The method according to any of the preceding claims, further comprising: e1) based on the synthesized data (24), training a machine learning algorithm for analyzing sensor data provided by a sensor device for perceiving a scene and / or environment thereof.

6. The method according to claim 5, characterized in that the sensor device is included in an autonomous vehicle for perceiving a scene and / or environment thereof and in that the method further comprises: e2) causing the autonomous vehicle to navigate based on the analysis.

7. The method according any of the preceding claims, further comprising one, several, or all of the following: f1) based on the synthesized data (24), training a three-dimensional objectdetection algorithm for detecting one or more objects and / or entities from sensor data provided by a sensor device for perceiving a scene and / or environment thereof; and / or f2) based on the synthesized data (24), training a depth estimation algorithm for estimating a depth of an object and / or entity from sensor data provided by a sensor device for perceiving a scene and / or environment thereof; and / orf3) based on the synthesized data (24), training an instance segmentation algorithm for distinguishing between instances of objects and / or entities of same category; and / or f4) based on the synthesized data (24), training a semantic segmentation algorithm for distinguishing between objects and / or entities of different categories; and / or f5) based on the synthesized data (24), training a panoptic segmentation algorithm for distinguishing between instances of objects and / or entities of same and / or different categories.

8. The method according to any of the preceding claims, one, several, or all of the following: g1) adapting the synthesized data (24) to one or more parameters of a sensor device for perceiving a scene and / or environment thereof; and / or g2) projecting the synthesized data (24) from three dimensions on two dimensions or vice versa; and / or g3) based on the synthesized data (24), creating three-dimensional point cloudbased data (38); and / or g4) based on the synthesized data (24), creating two-dimensional image data (40).

9. The method according to any of the preceding claims, characterized in that the synthetic depth data and / or reference depth data include: hi) one or more three-dimensional point variables for each of the synthetic objects and / or entities and / or of the reference objects and / or entities, respectively; and / or h2) a three-dimensional bounding box variable for each of the synthetic objects and / or entities and / or of the reference objects and / or entities, respectively.

10. The method according to claim 9, characterized in that the generating and / or discriminating includes predicting the one or more three-dimensional point variables and / or the three-dimensional bounding box variable of the synthetic depth data (26) and / or reference depth data (28), preferably by using a three- dimensional object detector algorithm.11 . The method according to any of the preceding claims, characterized in that the loss and / or gain function (16) is configured:11) as a L1 and / or L2 loss and / or gain function; and / or12) for monocular depth estimation.

12. A computing device comprising means adapted to execute the steps of the method according to any of the preceding claims.

13. A computer program comprising instructions to cause a computing device to execute the steps of the method according to any of the claims 1 to 11 .

14. A computer-readable storage medium having stored thereon the computer program of claim 13.

Images