WO2022043204A1 - Trainingsverfahren für einen generator zur erzeugung realistischer bilder - Google Patents
Trainingsverfahren für einen generator zur erzeugung realistischer bilder Download PDFInfo
- Publication number
- WO2022043204A1 WO2022043204A1 PCT/EP2021/073127 EP2021073127W WO2022043204A1 WO 2022043204 A1 WO2022043204 A1 WO 2022043204A1 EP 2021073127 W EP2021073127 W EP 2021073127W WO 2022043204 A1 WO2022043204 A1 WO 2022043204A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- image
- images
- training
- generated
- semantic
- Prior art date
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T11/00—2D [Two Dimensional] image generation
- G06T11/001—Texturing; Colouring; Generation of texture or colour
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0475—Generative networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T11/00—2D [Two Dimensional] image generation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T11/00—2D [Two Dimensional] image generation
- G06T11/60—Editing figures and text; Combining figures or text
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T9/00—Image coding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/776—Validation; Performance evaluation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/70—Labelling scene content, e.g. deriving syntactic or semantic representations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Definitions
- the present invention relates to the training of a realistic image generator, which in turn can be used to train image classifiers.
- Corresponding image classifiers have to be trained with training images that were recorded in a large number of traffic situations. Obtaining the training images is comparatively difficult and expensive. Traffic situations that rarely occur in reality can be so underrepresented in the data set with the training images that the image classifier cannot optimally learn how to classify them correctly. Furthermore, a lot of manual work is necessary to “label” the training images or their pixels with the associated target class assignments (“ground truth”).
- the concept of the image is not limited to static camera images, but also includes, for example, video images, radar images, lidar images and ultrasound images.
- the images to be generated can be realistic, for example in relation to a given application.
- “realistic” can mean in particular that the images can be used in downstream processing, such as when training an image classifier, in the same way as images recorded with physical sensors.
- realistically generated images can be used, for example, to enrich a stock of real training images recorded with sensors and then “labeled” for an image classifier.
- the images to be generated are therefore referred to below as “realistic images” or as “realistically generated images” in the interest of better legibility.
- the generator creates the realistic images from a semantic map.
- This semantic map assigns a semantic meaning of an object to which this pixel belongs to each pixel of the realistic image to be generated. So not just any random realistic image is generated, but one that reflects the situation given in the semantic map.
- the semantic map can designate a traffic situation with different lanes, lane boundaries, traffic signs, road users and other objects.
- Training maps that give each pixel of the respective training image a semantic assign meaning, provided. So there is a semantic training map for every real training image. Conversely, there is at least one real training image for each semantic training map because, for example, a semantically identical situation with different exposures or other imaging parameters may have been recorded.
- the semantic training maps can be obtained, for example, by manually labeling the real training images.
- a mixed image is generated from at least one realistic image generated by the generator and at least one real training image determined for the same semantic training map.
- a first real subset of the pixels is occupied by corresponding pixel values of the realistic image generated by the generator.
- the remaining real subset of pixels is occupied with corresponding pixel values of the real training image.
- Each pixel of the mixed image is therefore assigned either the corresponding pixel value of the realistic image generated by the generator or the corresponding pixel value of the real training image.
- contiguous areas of pixels of the mixed image can be assigned either uniformly with corresponding pixel values of the realistic image generated by the generator or uniformly with corresponding pixel values of the real training image.
- the mixed image can then be, for example, a “collage” of representations of objects in the realistic image generated by the generator on the one hand and representations of objects in the real training image on the other.
- the realistic images generated by the generator, the at least one real training image and the at least one mixed image are fed to the discriminator.
- Generator parameters which characterize the behavior of the generator, are optimized with the aim that the realistic images generated by the generator are misclassified as real images by the discriminator.
- discriminator parameters that characterize the behavior of the discriminator are optimized with the aim of improving the accuracy in distinguishing between realistically generated images and real images.
- the discriminator is thus trained to classify a realistically generated image as a realistically generated image and to classify a real training image as a real training image.
- the discriminator may be desirable for the discriminator to classify as a realistically generated image a mixed image that predominantly contains pixels and/or objects taken from the realistically generated image. It can also be desired, for example, for the discriminator to classify as a real image a mixed image that predominantly contains pixels and/or objects taken from the real training image. Any gradations in between are also possible.
- the parameters of the discriminator can therefore be optimized so that the discriminator outputs the respectively desired target assignment in response to the mixed image.
- the discriminator parameters can, for example, also be optimized with the aim of classifying the mixed image as a real image to an extent that corresponds to the numerical proportion of pixels and/or objects taken from a real training image into the mixed image. So if, for example, 60% of the image content of a mixed image from a real training image and 40% of the image content this mixed image was taken from a realistically generated image, it may be desirable for the discriminator to classify the mixed image as a real image with a score of 0.6 and as a realistically generated image with a score of 0.4.
- a PatchGAN discriminator for example, can be selected as the discriminator.
- a discriminator determines whether a realistically generated image or a real image is present in partial areas of the images with a predetermined size (“patches”). The results obtained in each case are then combined to form an overall result.
- Such a discriminator is particularly well able to quantitatively detect the mixing ratio of real to realistically generated image content in the mixed image.
- the discriminator can, for example, also have an encoder-decoder arrangement with an encoder structure and a decoder structure.
- the encoder structure translates an input image into an information-reduced representation in several processing layers.
- the decoder structure further translates this information-reduced representation into an assessment of each pixel of the input image as a real or realistically rendered pixel.
- the output of such a discriminator is not just a score that evaluates the input image as a whole. Instead, the evaluation is spatially resolved and can therefore also record in detail which pixels or Objects of the mixed image come from the real image and which pixels or objects of the mixed image come from the realistically generated image.
- the discriminator has at least one direct connection between a processing layer of the encoder structure and a processing layer of the decoder structure, bypassing the information-reduced representation. A particularly relevant portion of the information from the encoder structure can then be selectively transferred to the decoder structure without having to pass the "bottleneck" of the maximum information-reduced representation. This gives the discriminator a "U-Net" architecture.
- the discriminator is also trained to generate a spatially resolved output from a mixed image, which was determined according to a specified rule from a real training image and a realistically generated image, which is as close as possible to a mixture of the real training image on the one hand and for the realistically generated image on the other hand the outputs obtained according to the same predetermined rule.
- the discriminator is then equivariant under the merging of the images to the merged image.
- the scenery according to the semantic map has a vehicle in the top left corner of the image and a tree in the bottom right corner of the image.
- the specified rule states that the composite image should combine the vehicle extracted from the realistically generated image with the tree extracted from the real training image.
- the spatially resolved output determined by the discriminator for the mixed image will therefore classify the area with the vehicle as a realistically generated image portion and the area with the tree as a real image portion.
- the discriminator When the discriminator is applied to the real image, its spatially resolved output should fully classify that real image as a real image.
- its spatially resolved output of this realistically rendered image should be completely as classify realistically generated image. If these two spatially resolved outputs are now combined in the same way as the mixed image, the result should be that the top left corner is classified as the real image portion and the bottom right corner as the realistically generated image portion. This is the result that is also obtained when the mixed image is first formed and then the spatially resolved output is determined.
- the cost function (loss function) of the discriminator can be expanded by a consistency term L c of the form be expanded.
- D is the spatially resolved output of the discriminator and M denotes the operation of merging according to the given rule, x is the real image and x is the realistically generated image.
- L c The motivation of the consistency term L c is somewhat comparable to the consistency check when an unsorted amount of cash in coins and/or bills is counted by two different people. Then both people can start counting according to different schemes. For example, the first person may grab the coins and/or bills in random order and add up the values of each, while the second person forms packages of specified numbers of coins and/or bills of the same denomination and then adds the values of those packages. Both counting methods should end up with the same amount of money.
- the discriminator is required to respect the natural semantic class boundaries.
- the generated images are therefore not only realistic at the level of the individual pixels, but also take into account the shapes that the image areas assigned to the different object types have according to the semantic map.
- the spatially resolved output can include, for example, an output from the last layer of a neural network of the discriminator, from which the classification of the input image as real or realistic is generated and probabilities for both classifications emerge.
- the last layer can contain, for example, "logits", ie classification scores that have not yet been normalized with the Softmax function.
- an essential application of the training method described here is to enlarge a training data set for an image classifier and thus train the image classifier better overall, starting from a given training data set with real training images and associated target assignments to semantic meanings. Therefore, the invention also relates to a method for training an image classifier that assigns a semantic meaning to an input image, and/or pixels of this input image.
- a generator is trained according to the method described above. With this trained generator, realistic images are generated from semantic maps. These semantic maps are then no longer limited to those semantic maps that were used to train the generator, but can describe any desired scenarios.
- Semantic target meanings are determined from the semantic maps, onto which the trained image classifier is to map the realistic images in each case.
- the target meanings can include, for example, belonging to one or more classes of a predefined classification. For example, if a vehicle is drawn in at a specific location in the semantic map, then the realistically generated image will contain a vehicle at this location. Therefore, the image classifier should assign at least this image area to the “Vehicle” class.
- a training data set for the image classifier which contains real training images and associated semantic target meanings, is expanded to include the realistically generated images and associated semantic target meanings.
- the image classifier is trained with the extended training data set.
- the training data set can be enriched in this way, in particular, with realistic images of situations that were previously in the training data set were underrepresented. In this way, the image classifier can be better able to handle these situations.
- training images of rare but dangerous traffic situations are often difficult to obtain.
- fog, extreme snowfall or black ice which are part and parcel of the situation, may rarely be present.
- Other parts of the situation such as two vehicles on a collision course, may be too dangerous to replicate with real vehicles.
- the invention therefore also relates to a further method.
- an image classifier is trained as previously described using realistic images generated with the trained generator.
- images that were recorded with at least one sensor carried by a vehicle are assigned a semantic meaning.
- a control signal is determined from the semantic meaning determined by the image classifier. The vehicle is controlled with this control signal.
- the improved training advantageously improves the accuracy of the semantic meaning provided by the image classifier. Therefore, the probability that the reaction of the vehicle triggered by the control signal is appropriate to the traffic situation shown in the images is advantageously increased.
- the invention therefore also relates to a computer program with machine-readable instructions which, when executed on one or more computers, cause the computer or computers to carry out one of the methods described.
- control devices for vehicles and embedded systems for technical devices that are also able to execute machine-readable instructions are also to be regarded as computers.
- the invention also relates to a machine-readable data carrier and/or a download product with the computer program.
- a download product is a digital product that can be transmitted over a data network, ie can be downloaded by a user of the data network, and which can be offered for sale in an online shop for immediate download, for example.
- a computer can be equipped with the computer program, with the machine-readable data carrier or with the downloadable product.
- FIG. 1 exemplary embodiment of the method 100 for training the generator 1
- FIG. 2 illustration of the formation of a mixed image 6
- FIG. 3 exemplary embodiment of the method 200 for training the image classifier 9
- Figure 4 exemplary embodiment of the method 300 with a complete chain of effects up to the control of a vehicle 50.
- FIG. 1 is a schematic flowchart of an exemplary embodiment of the method 100.
- step 110 real training images 5 and associated semantic training maps 5a are provided.
- the semantic training maps 5a assign a semantic meaning 4 to each pixel of the respective training image 5.
- step 120 realistic images 3 are generated from at least one semantic training map 5a using the generator 1 to be trained.
- step 130 at least one real training image 5 is determined for the same at least one semantic training map 5a. For example, this can be that training image 5 through whose "labeling" the semantic training map 5a was created in the first place.
- a mixed image 6 is generated from at least one realistic image 3 generated by generator 1 and at least one real training image 5 determined in step 130.
- a first real subset 6a of the pixels is occupied by corresponding pixel values of the realistic image 3 generated by the generator 1.
- the remaining real subset 6b of the pixels is occupied with corresponding pixel values of the real training image 5 in each case.
- contiguous areas 61, 62 of pixels of the mixed image 6, to which the semantic training map 5a assigns the same semantic meaning 4 can be assigned either uniformly with corresponding pixel values of the realistic image 3 generated by generator 1 or uniformly with corresponding pixel values of the real training image 5 will.
- the formation of the mixed image 6 is illustrated in detail in FIG.
- step 150 the realistic images 3 generated by the generator 1, the at least one real training image 5 and at least one mixed image 6, all of which belong to the same semantic training map 5a, are fed to a discriminator 7.
- This discriminator 7 is designed to distinguish between realistic images 3 generated by the generator 1 and real images 5 of the scenery specified by the semantic training map 5a.
- the discriminator 7 is only required for training. When the completely trained generator 1 is later used, the discriminator 7 is no longer required.
- a PatchGAN discriminator can be selected as discriminator 7 .
- a PatchGAN discriminator determines the distinction between realistically generated images 3 and real images 5 in partial areas of the images 3, 5, 6 with a predetermined size and combines the results obtained in each case into an overall result.
- a discriminator 7 with an encoder-decoder arrangement can be selected.
- the encoder structure in this encoder-decoder arrangement translates an input image into an information-reduced representation in several successive processing layers.
- the decoder structure in the encoder-decoder arrangement further translates this information-reduced representation into an evaluation of each pixel of the input image as a real or realistically generated pixel.
- at least one direct connection between a processing layer of the encoder structure and a processing layer of the decoder structure can be provided in the discriminator 7, for example, bypassing the information-reduced representation.
- step 160 generator parameters la, which characterize the behavior of generator 1, are optimized with the aim that realistic images 3 generated by generator 1 are misclassified as real images 5 by discriminator 7.
- discriminator parameters 7a which characterize the behavior of discriminator 7, are optimized with the aim of improving the accuracy in distinguishing between realistically generated images 3 and real images 5.
- the optimization of the discriminator parameters 7a according to block 171 can also be aimed at the goal of classifying the mixed image 6 as a real image 5 to an extent (i.e. with a score) that corresponds to the numerical proportion of the results from a real training image 5 corresponds to pixels and/or objects taken over into the mixed image 6 .
- the discriminator 7 can also be trained so that it is determined from a mixed image 6, which is determined from a real training image 5 and a realistically generated image 3 according to a predetermined rule was generated, a spatially resolved output which is as close as possible to a mixture of the outputs obtained for the real training image 5 on the one hand and for the realistically generated image 3 on the other hand according to the same prescribed rule.
- the fully trained state of the generator parameters la is denoted by the reference symbol la*.
- the fully trained state of the discriminator parameters 7a is denoted by the reference symbol 7a*.
- FIG. 2 uses a simple example to illustrate how a mixed image 6 can be formed.
- a semantic map 2 is specified. This semantic map 2 assigns the semantic meaning 4 to a first area 21, that the images 3, 5 used are intended to show a book there. The semantic map 2 assigns the semantic meaning 4 to a second area 22, that the images 3, 5 used should show a table there.
- a realistic image 3 generated with the generator 1 shows, in accordance with the semantic map 2, a table 32 on which a book 31 lies.
- a real training image 5 shows another table 52 on which another book 51 is lying.
- the mixed image 6 is a collage of the table 32 in the realistic image 3 generated by the generator 1 and the book 51 in the real training image 5.
- FIG. 3 is a schematic flowchart of an exemplary embodiment of the method 200 for training an image classifier 9.
- a generator 1 is trained using the method 100 described above.
- realistic images 3 are generated from semantic maps 2 with the trained generator 1.
- the semantic maps 2 used in each case become in Step 230 determines semantic target meanings onto which the image classifier 9 is to map the realistic images 3 or pixels thereof.
- FIG. 4 is a schematic flow chart of an embodiment of the
- step 310 an image classifier 9 is trained using the method 200 described above. With this image classifier 9, in step 320, images 5 that were recorded with at least one sensor 50a carried by a vehicle 50 are assigned a semantic meaning 4. From this determined by the image classifier 9 semantic
- a control signal 330a is determined in step 330.
- the vehicle 50 is controlled with this control signal 330a.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Databases & Information Systems (AREA)
- Medical Informatics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Image Analysis (AREA)
Abstract
Description
Claims
Priority Applications (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US17/999,000 US20230177809A1 (en) | 2020-08-24 | 2021-08-20 | Training method for a generator for generating realistic images |
KR1020237010284A KR20230057434A (ko) | 2020-08-24 | 2021-08-20 | 사실적 이미지들의 생성을 위한 생성기를 위한 트레이닝 방법 |
JP2023513119A JP7505117B2 (ja) | 2020-08-24 | 2021-08-20 | 写実的画像を生成する生成器のためのトレーニング方法 |
CN202180051499.7A CN115989524A (zh) | 2020-08-24 | 2021-08-20 | 用于产生逼真图像的生成器的训练方法 |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
DE102020210710.6 | 2020-08-24 | ||
DE102020210710.6A DE102020210710A1 (de) | 2020-08-24 | 2020-08-24 | Trainingsverfahren für einen Generator zur Erzeugung realistischer Bilder |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2022043204A1 true WO2022043204A1 (de) | 2022-03-03 |
Family
ID=77543527
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/EP2021/073127 WO2022043204A1 (de) | 2020-08-24 | 2021-08-20 | Trainingsverfahren für einen generator zur erzeugung realistischer bilder |
Country Status (6)
Country | Link |
---|---|
US (1) | US20230177809A1 (de) |
JP (1) | JP7505117B2 (de) |
KR (1) | KR20230057434A (de) |
CN (1) | CN115989524A (de) |
DE (1) | DE102020210710A1 (de) |
WO (1) | WO2022043204A1 (de) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR20240165772A (ko) | 2023-05-16 | 2024-11-25 | 고려대학교 산학협력단 | 토큰 유사도 기반의 벡터-양자화 확산 모델을 활용한 고품질 이미지 생성 방법 및 이를 위한 장치 |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
DE102018204494B3 (de) | 2018-03-23 | 2019-08-14 | Robert Bosch Gmbh | Erzeugung synthetischer Radarsignale |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10678244B2 (en) * | 2017-03-23 | 2020-06-09 | Tesla, Inc. | Data synthesis for autonomous control systems |
US11899748B2 (en) * | 2019-09-06 | 2024-02-13 | Volkswagen Aktiengesellschaft | System, method, and apparatus for a neural network model for a vehicle |
-
2020
- 2020-08-24 DE DE102020210710.6A patent/DE102020210710A1/de active Pending
-
2021
- 2021-08-20 WO PCT/EP2021/073127 patent/WO2022043204A1/de active Application Filing
- 2021-08-20 JP JP2023513119A patent/JP7505117B2/ja active Active
- 2021-08-20 US US17/999,000 patent/US20230177809A1/en active Pending
- 2021-08-20 KR KR1020237010284A patent/KR20230057434A/ko active Pending
- 2021-08-20 CN CN202180051499.7A patent/CN115989524A/zh active Pending
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
DE102018204494B3 (de) | 2018-03-23 | 2019-08-14 | Robert Bosch Gmbh | Erzeugung synthetischer Radarsignale |
Non-Patent Citations (3)
Title |
---|
MICHAL URICAR ET AL: "Yes, we GAN: Applying Adversarial Techniques for Autonomous Driving", ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, 9 February 2019 (2019-02-09), XP081590898, DOI: 10.2352/ISSN.2470-1173.2019.15.AVM-048 * |
SCHONFELD EDGAR ET AL: "A U-Net Based Discriminator for Generative Adversarial Networks", 2020 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR), IEEE, 13 June 2020 (2020-06-13), pages 8204 - 8213, XP033803473, DOI: 10.1109/CVPR42600.2020.00823 * |
XIHUI LIU ET AL: "Learning to Predict Layout-to-image Conditional Convolutions for Semantic Image Synthesis", ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, 15 October 2019 (2019-10-15), XP081575876 * |
Also Published As
Publication number | Publication date |
---|---|
US20230177809A1 (en) | 2023-06-08 |
JP2023538444A (ja) | 2023-09-07 |
DE102020210710A1 (de) | 2022-02-24 |
KR20230057434A (ko) | 2023-04-28 |
JP7505117B2 (ja) | 2024-06-24 |
CN115989524A (zh) | 2023-04-18 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
DE102019209644A1 (de) | Verfahren zum Trainieren eines neuronalen Netzes | |
WO2020193510A1 (de) | Training für künstliche neuronale netzwerke mit besserer ausnutzung der lern-datensätze | |
WO2013152929A1 (de) | Lernverfahren zur automatisierten erkennung von verkehrszeichen, verfahren zur bestimmung eines aktualisierten parametersatzes für eine klassifikation von einem verkehrszeichen und verkehrszeichenerkennungssystem | |
DE102019214200A1 (de) | Übersetzung von Trainingsdaten zwischen Beobachtungsmodalitäten | |
WO2019206792A1 (de) | Verfahren und vorrichtung zur umsetzung eines eingangsbildes einer ersten domäne in ein ausgangsbild einer zweiten domäne | |
EP3850536A1 (de) | Analyse dynamisscher räumlicher szenarien | |
DE102020210711A1 (de) | Training eines Generators zur Erzeugung realistischer Bilder mit einem semantisch segmentierenden Diskriminator | |
DE102019208733A1 (de) | Verfahren und Generator zum Erzeugen von gestörten Eingangsdaten für ein neuronales Netz | |
DE102020208008A1 (de) | Bildklassifikation und zugehöriges Training für sicherheitsrelevante Klassifikationsaufgaben | |
WO2022043204A1 (de) | Trainingsverfahren für einen generator zur erzeugung realistischer bilder | |
WO2022043200A1 (de) | Erzeugung realistischer bilder aus vorgegebenen semantischen karten | |
DE102021205447A1 (de) | Datenanreicherung für das Trainieren von Bildklassifizierern | |
DE102018132627A1 (de) | Verfahren zum Erfassen einer Umgebung eines Kraftfahrzeugs mittels zeitlicher Fusion von Bildern durch ein künstliches neuronales Netz; Steuereinheit, Fahrerassistenzsystem; Computerprogrammprodukt | |
DE102022208083A1 (de) | Trainieren eines neuronalen Netzwerks mit Hilfe von Wissensgraphen | |
DE112022003908T5 (de) | Informationsverarbeitungsvorrichtung, informationsverarbeitungsverfahren, servervorrichtung, fahrzeugvorrichtung und informationsverarbeitungsprogramm | |
WO2021175783A1 (de) | Computerimplementiertes verfahren und system zum erzeugen synthetischer sensordaten und trainingsverfahren | |
EP4033452B1 (de) | Domänenunabhängiges training von bildklassifikatoren | |
EP3895415A1 (de) | Transfer von zusatzinformation zwischen kamerasystemen | |
DE102021214464B3 (de) | Bewertung von Generatoren zur Erzeugung realistischer Bilder | |
DE102019108722A1 (de) | Videoverarbeitung für maschinelles Lernen | |
DE102019103192A1 (de) | Verfahren zum Erzeugen von Trainingsdaten für ein digitales, lernfähiges Kamerasystem | |
DE102021104672A1 (de) | Erzeugen von kontrafaktischen Bildern für die Auswertung von Bildklassifikatoren | |
DE102022208564A1 (de) | Auswertung von Messdaten mit angepasster Auflösungssteigerung | |
DE102021208156A1 (de) | Bildklassifikator mit geringerem Erfordernis gelabelter Trainingsdaten | |
DE102021209805A1 (de) | Anomaliekontrolle für technische Systeme anhand der Kompatibilität von Bilddaten zu einer vorgegebenen Verteilung |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 21762729 Country of ref document: EP Kind code of ref document: A1 |
|
ENP | Entry into the national phase |
Ref document number: 2023513119 Country of ref document: JP Kind code of ref document: A |
|
ENP | Entry into the national phase |
Ref document number: 20237010284 Country of ref document: KR Kind code of ref document: A |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 21762729 Country of ref document: EP Kind code of ref document: A1 |