WO2018077745A1 - Procédé d'analyse pour marquages d'objet dans des images sur la base de modèles - Google Patents
Procédé d'analyse pour marquages d'objet dans des images sur la base de modèles Download PDFInfo
- Publication number
- WO2018077745A1 WO2018077745A1 PCT/EP2017/076847 EP2017076847W WO2018077745A1 WO 2018077745 A1 WO2018077745 A1 WO 2018077745A1 EP 2017076847 W EP2017076847 W EP 2017076847W WO 2018077745 A1 WO2018077745 A1 WO 2018077745A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- analysis method
- model
- object mark
- mark
- image
- Prior art date
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/20—Scenes; Scene-specific elements in augmented reality scenes
-
- 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
- G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
-
- 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
Definitions
- the invention relates to a method for analyzing object markers in images and to an analysis unit for carrying out the method.
- the present invention discloses an analysis method that allows to perform a quality control on image data with a label and to make a statement as to how much an object mark deviates in position from a position assumed to be ideal.
- the invention therefore provides an analysis method and an analysis unit according to the independent claims. Further developments of the invention are the subject of the dependent claims.
- an analysis method for checking a position of at least one object mark of at least one image showing a predetermined object reading at least one image with an object mark from a memory and checking at least once if the object mark is for the predetermined Object corresponds to at least one of at least two classes of a model, each class defining at least one position and / or size for the object mark with respect to the predetermined object, and wherein the model calculates a score for each test and outputs a signal based on that score, which indicates a quality of the object mark.
- the signal may indicate correctness of the position of the object mark for each image. It may indicate a deviation of the position and / or size of the object mark from an ideal position in one direction for each image, preferably indicating a measure of goodness and / or deviation in pixels. It may be a type and / or scope of an occurring error in the Display position and / or size of the object mark, in particular a deviation in one direction. It may indicate a distribution of at least one error in the position and / or size of the object marker. For each image, a position and / or size of the object mark may be specified, in particular a distance of the object mark from an edge of the object in the image.
- At least one descriptor may be defined which describes the at least one image and / or object, in particular at least one specific pattern which results, for example, from a displacement of the object marking.
- the descriptor may preferably be a HOG descriptor or a descriptor based on fast Fourier transform features and / or image gradients.
- Each model can be constructed for discrete shifts of the object mark in one direction with at least one classifier, preferably SVM-based.
- Each model may be evaluated with test data for determining a quality of a position and / or size of the object mark, wherein preferably the score indicates a correctness of the position and / or size of the object mark.
- Each model may have a first class corresponding to an ideal position of the object mark and having a second class corresponding to an unidirectionally shifted object mark.
- a first class of each model may correspond to an ideal position of the object marker, and a second class may correspond to a displacement associated with the model. If a score is determined to be above / below a predefined threshold for a model, the object marker can no longer be checked against any other models.
- the object mark can be a Model that corresponds to the displacement of the object mark.
- Each model may have a first class corresponding to an object mark having a first shift in one direction and a second class corresponding to an object mark having a second shift, in particular a greater shift in direction than the first class.
- the object mark can be evaluated with another model.
- the other model may be chosen depending on the score, and in the other model, the displacement of the first class object marker and the second class object marker displacement may be changed.
- the object mark can be tested against models that divide the space of possible displacements.
- the classes of a plurality of models may not substantially correspond to the ideal position of the object mark. Compared to the multitude few models can correspond to the ideal position.
- a model may have a class that corresponds to an ideal object marker and several classes that each correspond to a displacement of the object marker in one direction.
- the object mark can be tested against several classes of a model. Based on a score, the shift most likely to correspond to the actual shift can be determined.
- the at least one model can be created by means of training data with images and object markers. For each image, it can be specified to which object class an object shown in the picture belongs.
- the training data may be at least partially extracted from the images and object markers by defined displacement of the object markers at least a part of the images are generated.
- a model can be improved through online learning.
- the predetermined and / or shown object may be a pedestrian, cyclist, vehicle and / or object.
- the object mark can be tested against the model, preferably any model of a variety of models.
- a computer program product wherein the computer program product stores computer program instructions in computer readable memory, and the computer program instructions cause at least one analyzer to perform an analysis method as described herein when the computer program instructions are read and / or executed by the analyzer ,
- an analysis unit for verifying at least one object mark in image data comprising at least one memory storing the image data, data on object marks and / or the training data, and a processor adapted to perform an analysis process, such as it is described herein, wherein the image data, data on object marks and / or the training data are read from the memory by the analysis unit, the data and processed by the processor, and wherein the analysis unit outputs signals of the analysis method, preferably in the or another memory.
- Fig. 2 is a schematically illustrated rating by
- Fig. 5 is a schematic representation of the analysis unit.
- the driver assistance systems for this purpose have sensors and in particular optical sensors, such as cameras, to detect the vehicle environment and to detect objects in the vehicle environment. Examples of objects in the vehicle environment are other objects or road users such as persons and especially cyclists or pedestrians.
- a driver assistance system can recognize an object in the surroundings of the vehicle in good time and, for example, initiate a braking or evasive maneuver, it is necessary to recognize a road user as early as possible and to assess how critical behavior of the road user for the vehicle or the course of the journey is. It is possible that by using image analysis method and parts of the object in the environment of the vehicle can be detected and in particular body parts of a person.
- Driver assistance systems can be prepared for their use with methods of machine learning. These are based on the automatic recognition of patterns in data from images and the determination of models for Differentiation of data from different classes. This distinction can be made without prior knowledge (unsupervised) or with prior knowledge (supervised) about the classes to be distinguished.
- an assistance system can be equipped with models that specify how a driver assistance system has to evaluate image data and which object classes are to be detected by the driver assistance system or how the objects are to be detected. Training is in this case with a plurality of images, each having objects of the object classes to be recognized. For example, a driver assistance system can be trained or configured to recognize cyclists or pedestrians.
- the objects are usually described by an object marker, also known as a "bounding box.”
- the bounding box is a polygon line that surrounds the object to be recognized, and in particular, the polygon line is formed as a rectangle that encloses the object Accordingly, the object can be described by the object marking, whereby the object marking can be created automatically or by manually marking the object with the object marking, the so-called "labein”.
- the labein is done either by an expert or by a layman (engineer).
- a label thus contains the imaged object and is defined by the image coordinates of the polygon line surrounding the object.
- a quality control of labein or object markers can be done manually.
- the labels are then checked manually with respect to the images and a random check is made by visual matching of the match (position, size) from the label and object imaged in the image (see Su, H., Deng, J. and Li, F (2012) Crowdsourcing Annotations for Visual Object Detection, in: AAAI Human Computation Workshop, pp. 40-46.).
- a sample selection can be made in a simple way, for example, every xth image (eg every 5th, 10th or 15th image) or every image with an above-average number of objects can be selected.
- a random or complete quality control can be performed by cross-validation and by renewed and redundant manual lab by two experts or laymen.
- Such a shape then clings to an object shown in the picture.
- An evaluation of the label then takes place by evaluating the object marking with reference to the form generated by the algorithm with qualitative marks (eg areal label - Ground Truth (see Vittayakorn, S. and Hays, J. (201)), Quality Assessment for Crowdsourced Object Annotations, In: Hoey, J., McKenna, S. and Trucco, E., eds., Proceedings of the British Machine Vision Conference, BMVA Press, pp. 109.1 - 109.1 1.).
- qualitative marks eg areal label - Ground Truth (see Vittayakorn, S. and Hays, J. (201)), Quality Assessment for Crowdsourced Object Annotations, In: Hoey, J., McKenna, S. and Trucco, E., eds., Proceedings of the British Machine Vision Conference, BMVA Press, pp. 109.1 - 109.1 1.).
- the behavior and performance of editors can be estimated automatically by determining numerical sizes (e.g., time required to label an object).
- numerical sizes e.g., time required to label an object.
- the quality of the label can be inferred (see Sameki, M., Gurari, D. and Betke, M. (2015), Characterizing Image Segmentation Behavior of the Crowd, Collective Intelligence Conference, San Francisco.).
- the analysis method thus analyzes at least one image based on the image data of the image.
- An object such as a pedestrian or a cyclist, is described in the image data by a so-called object marker.
- the goal is to make a quantitative and / or qualitative statement about the position of an object mark in an image.
- a quality of object markers for image data sets can be evaluated in various ways. Based on the rating, the object markers can be improved.
- a quality of individual object markings of a data record can be evaluated by first producing a binary one by the analysis method Output for correctness of the position and / or size of the object mark is made or issued.
- a signal can be output that indicates whether the object mark is correctly or incorrectly placed and / or dimensioned.
- the ideal position is defined in particular in a training phase.
- the deviation is then present in at least one direction, preferably in an x and / or y direction in an image plane. It can also be a measure of a quality spent.
- an evaluation of the quality of the object markers of an entire image data record and / or of a subset of the image data record can be output.
- a statement about a type and / or extent of the occurring errors in the position and / or size of the object marking can be output by the analysis method, eg a systematic deviation by, for example, 5 pixels in the x and / or y direction.
- a quantitative statement about a distribution of an error in the positioning of the object marking can be output.
- An analysis of the causes of the error eg an analysis of certain intensity patterns, can lead to deviations, eg from 5 to 1 1 pixels in the x and / or y direction.
- an object flag identified by the analysis method can be output, the respective error being detected when a threshold for a property of the object marker has been exceeded (for example for a displacement).
- the object marks identified as defective can then be checked manually.
- a suggestion for correcting the position and / or size of the object mark can be output based thereon and / or the object mark can be automatically corrected.
- the analysis method preferably determines a deviation with pixel precision
- the object marking can, for example, be in its position and / or size are changed in the direction of the ideal position and / or an ideal size.
- the control of the quality of the object markings can be done in real time.
- an assistance function for an agent can be provided by means of the analysis method, which gives indications of incorrect object markings, makes suggestions for a correction and / or carries them out automatically.
- the analysis method is based on models generated prior to execution of the analysis method (e.g., pedestrian or cyclist object markers). Creating the models requires data from images and object markers. Each image includes an object marker and an indication of which class the object shown in the image belongs to (e.g., the "pedestrian” or "cyclist” class). Further, the position and size of each object mark (e.g., enclosing polygon line or enclosing rectangle line) defining the object in the image is defined.
- object mark e.g., enclosing polygon line or enclosing rectangle line
- the models can be created using preferably high-quality output data (in each case image and object marking).
- output data in each case image and object marking.
- data such as e.g. a defined distance of the object mark from an edge of the object in the image.
- Other data can also be generated by defined shifts of correct object markers in pictures.
- Output data are images with correct object markings with regard to position and / or size and images which were generated from these by object marks moved in a defined manner. These form training data for the analysis method.
- a descriptor is selected to describe the images.
- a descriptor can be used which detects / describes specific patterns, which can result, for example, from a displacement of the object marking, especially in the edge regions of the object marking.
- the descriptor may preferably be an HOG descriptor (HOG stands for "histogram of oriented gradients") Descriptor based on features of a fast Fourier transform (FFT) and / or image gradients are used.
- HOG descriptor HOG stands for "histogram of oriented gradients”
- FFT fast Fourier transform
- Models for discrete displacements in at least one direction can be compared with a classifier e.g. SVM-based (SVM stands for "support vector machine”) can be created and an optimization of parameters for creating the models can be done by means of cross-validation.
- SVM-based support vector machine
- a first class is defined for each model, which corresponds to a correct object marking. This means that the object mark of an object falls into this first class if it is correctly defined and corresponds to the ideal object mark.
- the object marker can then be "mapped" to the first class, and a second class corresponds to a moved object marker, that is, an object marker that is in the x- and / or y-direction relative to the ideal object marker, thus creating a two-class problem for each model
- An object marker is either optimal, or falls into the second class.
- the first class of the model matches the ("matched") ideal position of the object mark, while in the second class the values change for at least one direction of a shift, eg in the x and / or y direction, and in particular increase or A classifier pertaining to the model outputs a score, and if it is above or below a previously defined threshold, preferably no further models are tested and the tested object mark can be assigned to a model corresponding to the actual displacement
- the second class of this model matches the object mark being tested, and the rating of that model is given as the quality of the position and / or size of the object mark being tested.
- a first class is defined for each model, which corresponds to an object marking with a relatively smaller displacement in at least one direction.
- a second class corresponds to an object mark with relative to the first class of greater displacement, in which therefore preferably the shift relative to the first class in the x and / or y direction has been increased (eg with positive / negative values for (x, y): for the first class (5,10) versus (15,20) for the second class, where (x, y) indicates relative displacement in the x and / or y direction).
- a rating by a model is usually a further evaluation by another model.
- a shift of the object mark of the first class and / or a shift of the object mark of the second class can be changed and preferably incremented.
- the space of possible displacements is divided into intervals, eg with a first generated model with first class (5,10) and second class (10,15) and a second model with first class (10,15) and second class (15, 20).
- the models thus form a decision tree in a search space of possible displacements, where the models form the nodes of the decision tree and a model selects in each case a branch of the tree which is tracked for further evaluation.
- An object mark to be tested is here checked against models which divide the space of possible shifts into intervals between two possible shifts. Both classes of the majority of models do not match the ideal position and / or size of the object mark. Only a few models in the search space test directly against the ideal position and / or size.
- the search space in the space of possible shifts can be narrowed.
- the search in the room and thus the selection of the model for the next test is done by optimizing the rating.
- Based on a threshold for the rating can After a series of tests, the model most likely to match the actual displacement of the object mark is identified.
- a multi-class model with more than two classes can be used.
- a first class of a model then corresponds to the correct or correctly placed and / or dimensioned object mark.
- Several other classes correspond to displacements of the object mark by a certain value in one direction, in particular the x and / or y direction.
- An object mark to be tested is then tested against the multi-class model. By optimizing the score, the class most likely to match the actual displacement can be identified.
- processors may be used, e.g. at least one CPU (central processing unit) and / or GPU (graphics processing unit).
- Fig. 1 schematically illustrates a flow of the analysis method.
- a first step S1 at least one image is read out with an object mark from a store.
- a second step S2 it is checked whether the object marking for an object corresponds to at least one class of a model.
- a rating is performed by at least one model, and an output of a signal based on the rating.
- Fig. 2 schematically illustrates an evaluation of the models with test data (images and object marks) for determining the quality of the position of the object mark, with which a statement about the correctness of the position of the object mark can be made.
- the accuracy of the predicted displacements in at least one direction, e.g. in the x and / or y direction, the object markings of a data record can be specified in percent.
- An improvement of the models used can be done over time through online learning. For online training, e.g. An Extreme Learning Machine can be used as a classifier. In this case, the models have e.g.
- the implementation of quality control is based on the deposited models of the analysis procedure.
- Input data is the images to be tested and their associated object markers.
- Each object tag from the training data, which was also described with a descriptor, is tested against the previously provided models.
- the further procedure depends on the previously selected procedure for creating the models.
- Object marks evaluated as being faulty are further processed according to the measure of the quality of the evaluation. Faulty object marks can be displayed and edited manually and / or automatically. In this case, preferably only object markings can be displayed and reworked, in which the quality of the evaluation is less than a threshold value. Object markings that are evaluated as faulty can also be automatically corrected and presented to a processor for checking, since the analysis method has determined the deviation in the x and / or y direction. In particular, object markers in which the quality is greater than a threshold can be corrected automatically, since the analysis method has determined the deviation in the x and / or y direction.
- an object mark e.g. as follows:
- application-specific images and object markers are used to generate further training data for defined displacements in the x and / or y direction of object markers.
- a description of the image features for an extracted object mark takes place in a second step with a descriptor.
- the models corresponding to different displacements in the x and / or y direction are generated. From images to be tested with object markers, the object mark is extracted in each case and described by the same or a different descriptor. A test is made with the models from one or more of the above approaches.
- the analysis method can then output and / or display at least one of the following outputs as signals: a (binary) statement about a correctness of the position of the object mark for each image, a quantitative statement for the deviation of the position of the object mark from the ideal position of an object mark in x and y-direction for each label, indicating a measure of the quality of the statement, a statement of the nature and extent of occurring Error in the positioning of the object marking, eg systematic deviation by 5 pixels in the y-direction, a quantitative statement about distribution of the errors in the positioning of the object marking, and / or an analysis of the causes of given errors, eg certain intensity patterns leading to deviations from 5 to 1 1 pixels in the y-direction.
- the analysis method described here can be integrated into a label production tool.
- a generated object marking can be evaluated immediately after the (manual) generation with the tool.
- An object mark can, if the analysis method finds too great a deviation, e.g. be provided for an examination and / or a manual release.
- the analysis method can support agents with the initial marking with the object marking.
- the operator may preferably edit an object marker visualized on a display unit by the analysis method in real time, eg, by editing with an input device (mouse, stylus, keyboard, ).
- the analysis method can indicate the selection of the optimal object marking by, for example, color coding.
- hardware implementation is useful.
- a performance increase also allows use for real-time applications, eg in the field of automated driving. Integration into a production can take place, for example, as shown in FIG. 3.
- An editor creates an object mark for an object in step S20, while the generated model analysis method performs a real time evaluation of the created object mark.
- the analysis method can automatically output a correction proposal for the created object marking to the processor, propose a different object marking and / or in the selection of a support other / better object mark (step S21). This corresponds to an "online" procedure with immediate feedback to the processor.
- Images and object markers can be fed to the generated models analysis method in step 30.
- the analysis method may then perform an automatic correction of the object mark (step S31), output hints for a manual correction of the object mark (step S32), and / or output an analysis result on cause of errors and / or an error structure (step S33).
- the analysis method thus allows an automatic evaluation of the position quality of object markers of a data record by using methods from machine learning. It provides a method for deriving quantitative statements about positional errors of the object marks through a special generation of training data and its evaluation with machine learning, and allows output of a correction recommendation for the object marks in combination with the score.
- a consistent rating performance can be achieved in comparison with a human operator and a higher overall quality of the object markers.
- Systematic and random errors in object marking can be detected automatically, saving time and money.
- Adhoc quality control can be performed while creating an object marker, resulting in a simplification.
- a complete review of a dataset may be done instead of a random check to achieve more robust results.
- An evaluation of the object marks without prior knowledge of the nature, composition or cause of the defects is possible, whereby an independent assessment performance is achieved.
- models are created using the ideal object marks and the defined changed object marks and images.
- a "grid" of models is generated by which the object markers to be tested are evaluated.
- An object mark is tested as a test candidate in the first approach with all models and a score is determined for each model, ie in particular a consolidated evaluation of the test candidate.
- the value for the test candidate is given as the quality that is best, that is, the highest or lowest rating. This can also be viewed as brute force testing the object marker against the models.
- the models are designed in such a way that, in principle, there is always a distinction between an optimum and a deviation. It then calculates the consolidated score per test from test candidate to model.
- a review of the test candidate is done so that each model can also make a decision as to which model will be used next to test the test candidate.
- a selective definition of a model sequence takes place within the test procedure, whereby a total reduction of the test scope is achieved.
- a decision run is run through, at whose nodes or at points of the branching, in each case, the check is made with respect to a model.
- the root of the tree preferably corresponds to a check as to whether an optimal positioning of the object marking is present.
- a mean or the greatest possible deviation from the ideal position and / or size of the object mark are thereby defined by the training data, which have the deliberate displacements of the object markings.
- the mean deviation can be determined by an average between optimum and maximum deviation.
- the decision tree can then also be run depending on whether the test data belongs to the respective class or not. For each model, in particular a link to further models is stored and, depending on a rating determined on the respective model, the analysis method decides which model is selected next. In this case, the evaluation codes a possible deviation or a deviation checked by the model with respect to the defined object marking for the test candidate. Overall, this defines a search path through the decision tree, which is determined by the models passed through or by the various tests.
- a score is determined such that along the path through the decision tree there is an overall score that allows a statement about the quality of the candidate candidate.
- a two-layer decision tree or a tree with two levels is used. Here already in the root node a decision is made, and on the second level is already the output.
- a score for a test candidate is then output.
- the neural network is in turn trained with the training data, which both define an optimal placement of an object mark in an image, as well as describe a dedicated falsified location of object markers.
- the result of the decision tree decision or the output of the comparisons then leads to a statement about the absolute deviation of the test candidates.
- a mapping to a quality value can then take place depending on the application. It defines which result is considered to be particularly good or unfavorable.
- the analysis method can be carried out by an analysis unit which can process the images, for example in the form of a video signal, preferably in real time, store them and output the signals.
- the analysis unit can for this purpose use at least one processor, eg a "system on chip.”
- the one or more CPUs / GPUs may also execute program code, in particular the analysis method may be defined by program code
- the at least one CPU / GPU may allow parallelized execution of computational operations, such as computation Using Open GL or pasting overlay image data into image data
- An Image Signal Processor (ISP) can be used to render images and, for example, receive data from images and / or image sensors and increase quality, for example, through automatic white balance and / or nonvolatile memory which at least partially stores the image data and / or the training image data. From the at least one volatile and / or non-volatile memory, the image data, data on object markings and
- An output of the signals can then be acoustically and / or optically on a display unit.
- the information of the signals may be output to the at least one and / or another volatile and / or nonvolatile memory, e.g. in a file.
- the image data and / or the training image data can also be supplied to the analysis unit via an optical sensor.
- the analysis method may initially check if an object defined by the object marker is of a certain type. Specifically, the analysis method may determine an aspect ratio of the object mark, compare with at least an aspect ratio threshold, and output a signal indicating whether the object is of the particular type or not when the aspect ratio threshold is exceeded.
- the analysis method can determine a deviation of the object marking from an ideal position and / or size of the object marking, in particular with pixel precision.
- the analysis method may automatically correct the deviation and / or output a signal indicating how the object mark is to be adjusted to the ideal position and / or size.
- the analysis method can check whether the correct object has been marked with the object mark on the image or whether the object belongs to a certain object class, ie whether, for example, a pedestrian has been marked. The analysis method can then test whether the type or type of object, For example, a pedestrian whose object mark is to be checked, is actually present, or whether another object type is present, for example, a vehicle or a cyclist. This test can also be carried out independently of the analysis method.
- the analysis method can recognize this.
- the analysis method may be used to check the object type e.g. Check aspect ratios of the object marks, where a tolerance range or at least a threshold for an aspect ratio can be defined. If an aspect ratio is not within the tolerance range, i. a particular aspect ratio is in particular above or below the at least one threshold, a signal is output for the image. The signal then indicates that the object marker is marking a wrong object, or the object is not of the type to be checked, ie it does not belong to the correct object class.
- the analysis procedure can then be ended.
- Fig. 5 shows schematically an analysis unit AE.
- the analysis unit AE is input data D received by the analysis unit AE, stored in at least one memory S and evaluated by at least one processor P and processed.
- the analysis unit AE then outputs the signals S accordingly.
- the analysis method in the form of computer program instructions may be stored in the memory S, and the computer program instructions may be executed at least by the analysis unit AE for performing an analysis method.
- the computer program instructions can be kept as a computer program product.
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Computational Linguistics (AREA)
- Image Analysis (AREA)
Abstract
La présente invention concerne un procédé d'analyse pour vérifier la position d'au moins un marquage d'objet d'au moins une image sur laquelle est représenté un objet prédéterminé, le procédé d'analyse comprenant les étapes consistant à lire au moins une image présentant un marquage d'objet depuis une mémoire et à vérifier au moins une fois si la marquage d'objet pour l'objet prédéterminé correspond au moins à une parmi au moins deux catégories d'un modèle, chaque catégorie définissant au moins une position et/ou taille pour le marquage d'objet vis-à-vis de l'objet prédéterminé, et le modèle calculant pour chaque vérification une évaluation et émettant en fonction de cette évaluation un signal qui indique la qualité du marquage d'objet.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP17793883.4A EP3529744A1 (fr) | 2016-10-24 | 2017-10-20 | Procédé d'analyse pour marquages d'objet dans des images sur la base de modèles |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
DE102016220872.1A DE102016220872A1 (de) | 2016-10-24 | 2016-10-24 | Analyseverfahren für Objektmarkierungen in Bildern auf Basis von Modellen |
DE102016220872.1 | 2016-10-24 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2018077745A1 true WO2018077745A1 (fr) | 2018-05-03 |
Family
ID=60245067
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/EP2017/076847 WO2018077745A1 (fr) | 2016-10-24 | 2017-10-20 | Procédé d'analyse pour marquages d'objet dans des images sur la base de modèles |
Country Status (3)
Country | Link |
---|---|
EP (1) | EP3529744A1 (fr) |
DE (1) | DE102016220872A1 (fr) |
WO (1) | WO2018077745A1 (fr) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110363304A (zh) * | 2019-06-18 | 2019-10-22 | 深圳壹账通智能科技有限公司 | 定位模型构建方法、装置、计算机设备和存储介质 |
CN110781908A (zh) * | 2018-07-24 | 2020-02-11 | 罗伯特·博世有限公司 | 传感器数据中的改善的对象标记 |
CN112183456A (zh) * | 2020-10-19 | 2021-01-05 | 北京深睿博联科技有限责任公司 | 基于样本生成和域适应的多场景运动目标检测方法及装置 |
DE102019124252A1 (de) * | 2019-09-10 | 2021-03-11 | Bayerische Motoren Werke Aktiengesellschaft | Verfahren zum Betreiben eines Fahrzeugs, Fahrzeug, Computerprogramm und computerlesbares Speichermedium |
DE102020107776A1 (de) | 2020-03-20 | 2021-09-23 | Bayerische Motoren Werke Aktiengesellschaft | Trainieren eines automatischen Erkennungssystems |
-
2016
- 2016-10-24 DE DE102016220872.1A patent/DE102016220872A1/de not_active Ceased
-
2017
- 2017-10-20 EP EP17793883.4A patent/EP3529744A1/fr active Pending
- 2017-10-20 WO PCT/EP2017/076847 patent/WO2018077745A1/fr active Application Filing
Non-Patent Citations (7)
Title |
---|
HAO SU ET AL: "Crowdsourcing Annotations for Visual Object Detection", 1 January 2012 (2012-01-01), XP055440011, Retrieved from the Internet <URL:http://vision.stanford.edu/pdf/bbox_submission.pdf> [retrieved on 20180111] * |
MEHRNOOSH SAMEKI ET AL: "Characterizing Image Segmentation Behavior of the Crowd", 1 January 2015 (2015-01-01), XP055440023, Retrieved from the Internet <URL:https://sites.lsa.umich.edu/collectiveintelligence/wp-content/uploads/sites/176/2015/02/Sameki-CI-2015-Abstract.pdf> [retrieved on 20180111] * |
SAMEKI, M.; GURARI, D.; BETKE, M.: "Characterizing Image Segmentation Behavior of the Crowd", COLLECTIVE INTELLIGENCE CONFERENCE, 2015 |
SIRION VITTAYAKORN ET AL: "Quality Assessment for Crowdsourced Object Annotations", PROCEDINGS OF THE BRITISH MACHINE VISION CONFERENCE 2011, 9 March 2011 (2011-03-09), pages 109.1 - 109.11, XP055440018, ISBN: 978-1-901725-43-8, DOI: 10.5244/C.25.109 * |
SU, H.; DENG, J.; LI, F.-F.: "Crowdsourcing Annotations for Visual Object Detection", AAAI HUMAN COMPUTATION WORKSHOP, 2012, pages 40 - 46 |
TABORSKY EMMA ET AL: "Annotating Unconstrained Face Imagery: A scalable approach", 2015 INTERNATIONAL CONFERENCE ON BIOMETRICS (ICB), IEEE, 19 May 2015 (2015-05-19), pages 264 - 271, XP033166156, DOI: 10.1109/ICB.2015.7139094 * |
VITTAYAKORN, S.; HAYS, J.: "Proceedings of the British Machine Vision Conference", 2011, BMVA PRESS, article "Quality Assessment for Crowdsourced Object Annotations", pages: 109.1 - 109.11 |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110781908A (zh) * | 2018-07-24 | 2020-02-11 | 罗伯特·博世有限公司 | 传感器数据中的改善的对象标记 |
CN110363304A (zh) * | 2019-06-18 | 2019-10-22 | 深圳壹账通智能科技有限公司 | 定位模型构建方法、装置、计算机设备和存储介质 |
DE102019124252A1 (de) * | 2019-09-10 | 2021-03-11 | Bayerische Motoren Werke Aktiengesellschaft | Verfahren zum Betreiben eines Fahrzeugs, Fahrzeug, Computerprogramm und computerlesbares Speichermedium |
DE102020107776A1 (de) | 2020-03-20 | 2021-09-23 | Bayerische Motoren Werke Aktiengesellschaft | Trainieren eines automatischen Erkennungssystems |
CN112183456A (zh) * | 2020-10-19 | 2021-01-05 | 北京深睿博联科技有限责任公司 | 基于样本生成和域适应的多场景运动目标检测方法及装置 |
CN112183456B (zh) * | 2020-10-19 | 2021-07-02 | 北京深睿博联科技有限责任公司 | 基于样本生成和域适应的多场景运动目标检测方法及装置 |
Also Published As
Publication number | Publication date |
---|---|
EP3529744A1 (fr) | 2019-08-28 |
DE102016220872A1 (de) | 2018-04-26 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2018077745A1 (fr) | Procédé d'analyse pour marquages d'objet dans des images sur la base de modèles | |
DE102018128158A1 (de) | Vorrichtung zur inspektion des erscheinungsbilds | |
DE102010013943B4 (de) | Verfahren und Vorrichtung für eine Funktionsprüfung einer Objekt-Erkennungseinrichtung eines Kraftwagens | |
DE102018119682A1 (de) | Bildverarbeitungsvorrichtung, bildverarbeitungsverfahren und nicht-vorübergehendes computerlesbares speichermedium | |
EP3393875B1 (fr) | Procédé de détection améliorée d'objets au moyen d'un système d'aide à la conduite | |
EP3948688B1 (fr) | Entraînement pour réseaux neuronaux artificiels avec une meilleure utilisation des jeux de données d'apprentissage | |
DE102020126554A1 (de) | Mikroskopiesystem und verfahren zum überprüfen von eingabedaten | |
DE112019002848T5 (de) | System und verfahren zum auffinden und klassifizieren von mustern in einem bild mit einem bildverarbeitungssystem | |
EP2546778A2 (fr) | Procédé destiné à l'évaluation d'un dispositif d'identification dýobjet dýun véhicule automobile | |
DE102019208735B4 (de) | Verfahren zum Betreiben eines Fahrassistenzsystems eines Fahrzeugs und Fahrerassistenzsystem für ein Fahrzeug | |
WO2019119011A1 (fr) | Modèle de comportement d'un capteur d'environnement | |
DE102020211636A1 (de) | Verfahren und Vorrichtung zum Bereitstellen von Daten zum Erstellen einer digitalen Karte | |
EP3983936A1 (fr) | Procédé et générateur servant à créer des données d'entrée perturbées destinées à un réseau neuronal | |
DE102019120696A1 (de) | Vorrichtung und Verfahren zur Reifenprüfung | |
DE102021114287A1 (de) | Mikroskopiesystem und Verfahren zum Erzeugen stilisierter Kontrastbilder | |
DE102018113621A1 (de) | Verfahren zum Trainieren eines konvolutionellen neuronalen Netzwerks zum Verarbeiten von Bilddaten zur Anwendung in einem Fahrunterstützungssystem | |
WO2018077746A1 (fr) | Procédé d'analyse pour marquages d'objet dans des images | |
DE112022002323T5 (de) | Erkennungsvorrichtung, steuerverfahren für erkennungsvorrichtung, verfahren zur modellerzeugung durch modellerzeugungsvorrichtung, die ein trainiertes modell erzeugt, informationsverarbeitungsprogramm und aufzeichnungsmedium | |
DE102021133868A1 (de) | Mikroskopiesystem und Verfahren zum Modifizieren von Mikroskopbildern im Merkmalsraum eines generativen Netzes | |
DE102021129864A1 (de) | Verfahren und System zur Annotation von Sensordaten | |
EP3772017A1 (fr) | Détection de signal ferroviaire pour véhicules ferroviaires autonomes | |
DE102019220615A1 (de) | Verfahren und Vorrichtung zum Erkennen und Klassifizieren von Objekten | |
DE102021206377A1 (de) | Robusteres überwachtes Training für Bildklassifikatoren | |
DE102022204364A1 (de) | Qualitätsbestimmung einer Objekterkennung eines neuronalen Netzes | |
DE102020126291A1 (de) | Verfahren zum Analysieren eines Bauteils, Verfahren zum Trainieren eines Systems, Vorrichtung, Computerprogramm und computerlesbares Speichermedium |
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: 17793883 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2017793883 Country of ref document: EP |