WO2016157499A1 - 画像処理装置、物体検知装置、画像処理方法 - Google Patents
画像処理装置、物体検知装置、画像処理方法 Download PDFInfo
- Publication number
- WO2016157499A1 WO2016157499A1 PCT/JP2015/060482 JP2015060482W WO2016157499A1 WO 2016157499 A1 WO2016157499 A1 WO 2016157499A1 JP 2015060482 W JP2015060482 W JP 2015060482W WO 2016157499 A1 WO2016157499 A1 WO 2016157499A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- image processing
- unit
- classifier
- processing apparatus
- image
- Prior art date
Links
Images
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
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
- G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
- G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
- G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
- G06V10/454—Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/2431—Multiple classes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/70—Determining position or orientation of objects or cameras
-
- 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
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2200/00—Indexing scheme for image data processing or generation, in general
- G06T2200/04—Indexing scheme for image data processing or generation, in general involving 3D image data
-
- 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
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30248—Vehicle exterior or interior
- G06T2207/30252—Vehicle exterior; Vicinity of vehicle
Definitions
- the present invention relates to an image processing device, an object detection device, and an image processing method.
- object recognition technology based on machine learning is widely used.
- an in-vehicle preventive safety system that prevents traffic accidents and a monitoring system that notifies intrusion of a suspicious person have been put into practical use.
- learning and maintaining teacher images of objects to be recognized learning of classifiers that perform object recognition using machine learning techniques represented by support vector machines, boosting, multilayer neural networks, etc. It is carried out.
- Patent Document 1 The technology described in Patent Document 1 is known for improving the efficiency of maintenance of teacher data used for machine learning.
- a learning image of a three-dimensional shape model viewed from an arbitrary viewpoint is generated using computer graphics (CG), and a classifier is generated and learned using the generated learning image.
- CG computer graphics
- Patent Document 1 an image for learning of an arbitrary viewpoint can be generated, but an appropriate evaluation is made on the level of discrimination performance of the classifier generated and learned using the image for learning I can't.
- An image processing apparatus is for evaluating a classifier that identifies an object in an input image and classifies the object into one of a plurality of types, and uses the classifier And identifying each of the objects included in a plurality of verification images each having a known type of the object, and outputting any of the plurality of types for each of the verification images.
- An identification unit for obtaining identification performance; and an evaluation unit for outputting an evaluation result for the classifier based on the identification performance of the classifier obtained by the identification unit.
- An object detection device uses an identifier that has been learned using the above-described image processing device, detects an object in an image input from a camera, and detects the object using the object detection unit.
- An image processing method uses a computer for identifying an object in an input image and evaluating a discriminator that classifies the object into one of a plurality of types, The computer uses the classifier to identify the target object included in each of the plurality of verification images whose types of the target object are known, and determines any of the plurality of types for each verification image. By outputting, the discrimination performance of the discriminator is obtained, and the computer outputs the evaluation result for the discriminator based on the obtained discrimination performance of the discriminator.
- FIG. 1 is a block diagram showing a configuration of an image processing apparatus 10 according to the first embodiment of the present invention.
- An image processing apparatus 10 illustrated in FIG. 1 includes an input unit 101, a learning unit 102, an identification unit 103, an evaluation unit 104, a teacher data generation unit 105, and an output unit 106.
- the image processing apparatus 10 is connected to a database 107 storing verification data and a discriminator 108. Note that the database 107 and the identifier 108 may be provided inside the image processing apparatus 10, respectively.
- Each unit of the image processing apparatus 10 may be configured by hardware, or may be configured by software executed by a computer. Further, it may be a module combining hardware and software.
- the input unit 101 is a part for setting input data for the learning unit 102.
- the learning unit 102 is a part that learns the classifier 108 using data input from the input unit 101.
- the identification unit 103 is a part that identifies an object from verification data stored in the database 107 using the classifier 108 and obtains the classification performance of the classifier 108 based on the identification result.
- the evaluation unit 104 is a part that evaluates the discriminator 108 based on the discrimination performance of the discriminator 108 obtained by the discrimination unit 103.
- the teacher data generation unit 105 is a part that generates teacher data for the learning unit 102 to use for learning of the classifier 108 based on the evaluation result of the classifier 108 by the evaluation unit 104.
- the output unit 106 is a part that determines the learning status of the classifier 108 by the learning unit 102 and outputs the learned classifier 108.
- the identifier 108 identifies an object in the input image and classifies the object into one of a plurality of types.
- the discrimination performance of the discriminator 108 can be improved by learning performed by the learning unit 102.
- the verification data stored in the database 107 includes images of objects to be identified by the discriminator 108, a plurality of verification images whose types of objects are known, and types of objects in the respective verification images. And type information indicating what is. Although it is preferable to use a real image as the verification image, a CG image may be included.
- FIG. 2 is a flowchart for explaining the operation of the image processing apparatus 10 according to the first embodiment of the present invention.
- step S2001 the input unit 101 receives the input of the classifier 108 to be evaluated and learned, and the learning unit 102 uses it in learning of the classifier 108 based on the teacher data output from the teacher data generation unit 105. Set teacher data and learning parameters. Then, the input discriminator 108 and the set data are output as input data to the learning unit 102.
- step S2002 the learning unit 102 learns the discriminator 108 based on the teacher data and the learning parameters included in the input data from the input unit 101.
- the discriminator 108 for example, a multi-class discriminator represented by Deep Convolutional Neural Network (DCNN) can be used.
- a learning algorithm used by the learning unit 102 for learning of the discriminator 108 for example, an optimization algorithm such as a steepest descent method, a Newton method, a stochastic gradient descent method (SGD: Stochastic Gradient Descent) can be used.
- SGD stochastic gradient descent method
- a learning rate various activation functions (for example, sigmoid function, ReLU, hyperbola function, etc.), batch size, filter size, number of filters, etc.
- the discriminator 108 is not limited to DCNN.
- the discriminator 108 may be configured using a deep-neutral network (DNN) that is a fully connected network, a multi-class support vector machine, a logistic regression, or the like.
- DNN deep-neutral network
- the identification unit 103 uses the verification data stored in the database 107 to evaluate the identification performance of the classifier 108 that has learned in step S2002.
- the identification unit 103 identifies the objects included in each of the plurality of verification images in the verification data using the classifier 108, and the object is selected from any of a plurality of predetermined types. Classify. Then, by verifying the classification result of the object for each obtained verification image with the classification information of the verification data, it is verified whether the identification result of the object by the classifier 108 is correct. Judge every. By summing up the determination results, the identification unit 103 can obtain the identification performance of the classifier 108.
- FIG. 3 is a diagram illustrating an example of a table representing the discrimination performance of the discriminator 108 obtained by the discriminating unit 103.
- FIG. 3 shows an evaluation table 301 in which the identification results of the object with respect to the verification image are tabulated for each higher category and subcategory as an example of the identification performance of the classifier 108 in a table format.
- the upper category represents each class type in multi-class identification performed by the classifier 108.
- the discriminator 108 identifies various subjects in an image obtained by photographing the periphery of a vehicle as an object
- large cars, medium-sized cars, small cars, motorcycles, bicycles, pedestrians, backgrounds, and the like are high-level categories. Is set.
- the subcategory represents an attribute obtained by further subdividing each upper category.
- vehicle types such as sedans, light cars, trucks, and buses, body colors such as red, white, and black, object postures viewed from cameras such as 0 degrees, 45 degrees, and 90 degrees
- object postures viewed from cameras such as 0 degrees, 45 degrees, and 90 degrees
- background types such as urban areas, suburbs, and mountainous areas are set as subcategories.
- the identification unit 103 calculates evaluation values according to various evaluation criteria for each of these upper categories and subcategories based on the identification result of the object by the classifier 108 for each verification image. Then, the evaluation table 301 can be created by aggregating the calculated evaluation values.
- FIG. 3 shows an example in which N types of evaluation criteria are set, and evaluation values for each evaluation criterion are summarized in the evaluation table 301. Note that various values can be used as the evaluation value as an index for defining the discrimination performance of the discriminator 108. For example, the object detection performance, that is, the recall rate (Recall), which is an index for the comprehensiveness when detecting the object to be identified from the image, and the object detection accuracy, that is, the object to be identified from the image is detected.
- the recall rate recall
- An accuracy value (Precision) that is an index for accuracy, an F value (F-measure) that is an index considering both detection performance and detection accuracy, and the like can be used as evaluation values. Further, the maximum certainty factor, the minimum certainty factor, the average certainty factor, or the like can be used as the evaluation value for the classification result for each subcategory.
- step S2004 the evaluation unit 104 evaluates the identification error characteristic of the classifier 108 based on the identification performance evaluation result by the identification unit 103 in step S2003.
- the evaluation unit 104 lacks teacher data for at least one of the upper categories and at least one of the subcategories based on the evaluation values tabulated in the evaluation table 301 illustrated in FIG. Identify as a category.
- a combination of a higher category and a subcategory having a recall rate (Recall) equal to or less than a predetermined value, or a combination of a category and a subcategory with a negative average certainty is specified as a category for which teacher data is insufficient. .
- the evaluation unit 104 can determine the classification error characteristic of the classifier 108 by determining in which category the teacher data is insufficient in the evaluation unit 104 from the evaluation result of the identification performance by the identification unit 103. . Thereafter, the evaluation unit 104 outputs information on the upper category and the subcategory identified as the category for which the teacher data is insufficient as an evaluation result for the classifier 108 to the teacher data generation unit 105.
- the evaluation unit 104 may output information on subcategories other than those specified as described above based on the correlation between the subcategories.
- RV and SUV are set as subcategories representing vehicle types.
- the vehicles corresponding to these vehicle types have similar shapes and are considered to be highly correlated with each other. Therefore, when any one of these subcategories is specified, information on the other subcategory may also be output.
- step S2005 the teacher data generation unit 105 generates teacher data based on the category information output from the evaluation unit 104 as the evaluation result for the discriminator 108 in step S2004. Note that details of the method by which the teacher data generation unit 105 generates teacher data will be described later with reference to FIGS. 4, 5, 6, and 7.
- step S2006 the output unit 106 determines whether the learning status of the classifier 108 by the learning unit 102 satisfies a predetermined end condition.
- the learning status of the discriminator 108 is determined based on the evaluation result of the discriminating performance obtained by the discriminating unit 103 in step S2003. For example, when all the evaluation values in the evaluation table 301 illustrated in FIG. 3 are equal to or greater than a predetermined value, or when the cost reduction amount by learning is less than the predetermined value, the learning status of the discriminator 108 is an end condition. It can be determined that As a result, if it is determined that the learning status of the classifier 108 satisfies the termination condition, the process in FIG. 2 proceeds to step S2007. On the other hand, if it is determined that the learning status of the classifier 108 does not satisfy the termination condition, the process returns to step S2001 and the above-described processing is repeated.
- step S2006 the output unit 106 calculates an index value representing the learning status of the discriminator 108 from the evaluation value of each category in the evaluation table 301, and compares the index value with a predetermined reference value. Alternatively, it may be determined whether or not the learning status of the classifier 108 satisfies the termination condition. Specific processing contents in this case will be described below with reference to FIG.
- FIG. 4 is a diagram illustrating a change in the undetected rate for each category calculated based on the above-described recall rate (Recall) among the evaluated values in the evaluation table 301.
- the solid line 302 represents the state of the undetected rate Vu for each category obtained by the first learning
- the broken line 303 represents the undetected rate Vu for each category obtained by the second learning. It represents the situation.
- a one-dot chain line 304 represents a state of the undetected rate Vu for each category obtained by the N-th learning. As shown in FIG.
- the solid line 302 and the broken line 303 have a section (category) that exceeds the threshold Vth, whereas the alternate long and short dash line 304 is below the threshold Vth in all sections (categories).
- the undetected rate Vu is used as an index value of the learning status of the discriminator 108, and the end condition is that the undetected rate Vu is equal to or less than the threshold value Vth. It can be determined that the situation satisfies the termination condition.
- step S2007 the output unit 106 outputs the learned discriminator 108 obtained by the learning so far. Thereby, in the image processing apparatus 10, the learning of the discriminator 108 by the learning unit 102 ends.
- step S2007 When the process of step S2007 is executed, the image processing apparatus 10 completes the process shown in the flowchart of FIG.
- FIG. 5 is a block diagram illustrating a configuration example of the teacher data generation unit 105.
- the teacher data generation unit 105 illustrated in FIG. 5 includes object data 401, background data 402, parameter history information 403, an object setting unit 404, a background setting unit 405, a parameter setting unit 406, a teacher image generation unit 407, and an annotation unit 408.
- the object setting unit 404, the background setting unit 405, and the parameter setting unit 406 are collectively shown as the setting unit 40.
- the object data 401 is data representing a three-dimensional shape model of an object to be identified by the classifier 108, and includes geometric information and material information.
- the geometric information includes information on the shape of an object such as a point, line, or surface, information on a portion of a structured object such as a headlight, license plate, or tire.
- the material information is information regarding material properties of the object such as reflection, transmission, refraction, and light emission.
- the background data 402 is a three-dimensional shape model of the background to be identified by the classifier 108. Assuming image-based triting (IBL), the background data 402 may be held as a global image instead of a three-dimensional shape model. In this case, an image very close to a real image can be generated by combining a global image and a physical shader.
- the parameter history information 403 is parameter history information used when generating teacher data in the past.
- the object setting unit 404 sets object data used for generating a teacher image in the object data 401.
- the background setting unit 405 sets the background data used for generating the teacher image in the background data 402.
- the parameter setting unit 406 performs parameter setting for generating a teacher image.
- the teacher image generation unit 407 generates a teacher image based on the setting result in the setting unit 40.
- the annotation unit 408 generates teacher data based on the teacher image generated by the teacher image generation unit 407 and outputs the teacher data.
- FIG. 6 is a flowchart for explaining the operation of the teacher data generation unit 105.
- step S5001 the setting unit 40 receives the upper category and subcategory information output from the evaluation unit 104 in step S2004 of FIG. As described above, the evaluation unit 104 outputs information on the upper category and subcategory identified as the category for which the teacher data is insufficient as the evaluation result of the identification error characteristic of the classifier 108. In step S5001, this information is input to the object setting unit 404, the background setting unit 405, and the parameter setting unit 406 of the setting unit 40, respectively.
- the object setting unit 404 sets object data for generating a teacher image based on the information of the upper category and the subcategory received in step S5001.
- the object type such as “vehicle type: coupe”, “vehicle body color: black”, for example, is known as the category for which the teacher data is insufficient. Therefore, object data corresponding to such an object type is acquired from the object data 401 and set as object data for generating a teacher image.
- the color if the same color information does not exist in the object data 401, the color information may be replaced to generate an approximate color or a new color and set the object data.
- the object data set by the object setting unit 404 is not limited to data related to a single object, but may be data related to a plurality of types of objects.
- the background setting unit 405 sets background data for generating a teacher image based on the upper category and subcategory information received in step S5001.
- background types such as “urban area” and “western direct sunlight” are known as categories for which teacher data is insufficient. Therefore, background data relating to such a background type is acquired from the background data 402 and set as background data for generating a teacher image.
- the background data set by the background setting unit 405 is not limited to data related to a single background, but may be data related to multiple types of backgrounds.
- step S5004 the parameter setting unit 406 sets parameters for generating a teacher image based on the upper category and subcategory information received in step S5001.
- image generation parameters such as “45 degrees oblique as viewed from the camera” and “imaging distance is 50 m away” are known as categories for which the teacher data is insufficient. Therefore, such an image generation parameter is set as a parameter for generating a teacher image.
- a plurality of types of parameters may be set.
- the parameter setting unit 406 refers to the parameter history information 403 and does not set parameters that are the same as previously used parameters based on the contents.
- the teacher data generation unit 105 can generate teacher data using parameters that are different from parameters that have been used in the past.
- the teacher image generation unit 407 generates a teacher image based on the object data, background data, and parameters set by the object setting unit 404, the background setting unit 405, and the parameter setting unit 406 in steps S5002 to S5004, respectively.
- a teacher image is generated by combining object data and background data so that reflection, refraction, transmission, shadow, etc. are physically reproduced correctly.
- a teacher image may be generated in consideration of a physical camera. Thereby, in addition to basic camera internal parameters such as focal length and angle of view, a lens aberration can be taken into consideration to generate a teacher image.
- a plurality of types of object data, background data, and parameters may be set in steps S5002 to S5004.
- a plurality of teacher images are also generated according to the number of each setting. For example, when five types of object data, background data, and parameters are set, the teacher image generation unit 407 generates a total of 125 teacher images.
- step S5006 the annotation unit 408 generates correct data for the teacher image generated by the teacher image generation unit 407 in step S5005.
- the annotation unit 408 Based on the object data, background data, and parameters used to generate the teacher data, which part of the object to be identified is in the generated teacher image, which upper category and subcategory the object belongs to , Etc. Based on the determination result, the annotation unit 408 can generate correct answer data.
- the teacher data generation unit 105 When the process of step S5006 is executed, the teacher data generation unit 105 outputs the obtained teacher image and correct answer data as teacher data, and completes the process shown in the flowchart of FIG.
- FIG. 7 is a diagram for explaining a teacher image generation method.
- cameras 602, 603, and 604 are installed on a recognition target object 601 that is a vehicle, and a CG image corresponding to each captured image obtained by photographing the recognition target object 601 with these cameras is a teacher image. Shows the situation to generate.
- the recognition object 601 has three-axis rotation / translation parameters and three-axis scaling parameters.
- the cameras 602, 603, and 604 also have three-axis rotation / translation parameters, three-axis scaling parameters, internal parameters (focal length, image sensor size, image principal point, etc.), and lens aberration parameters (distortion coefficients). , Image height function, etc.). These parameters are set by the parameter setting unit 406.
- FIG. 8 is a diagram for explaining a method of generating correct data.
- a CG image corresponding to a photographed image obtained by photographing the recognition object 601 from the camera 603 in the situation of FIG. 7 is shown as an example of a teacher image.
- an object image 701 in which the recognition target object 601 in FIG. 7 is represented by a CG image and a background image 702 in which the road surface of the background portion is represented by a CG image are combined.
- bounding boxes 703, 704, 705, and 706 corresponding to the entire portion of the recognition target object 601 that is a vehicle, the license plate portion, the right headlight portion, and the left headlight portion are set. Yes.
- the position of the object image 701 in the generated teacher image can be calculated by a perspective projection method.
- the bounding box 703 can be calculated by perspectively projecting the entire three-dimensional shape information of the recognition target 601 represented by the object data on the image.
- the bounding boxes 704, 705, and 706 can be calculated in the same manner as the bounding box 703 by structuring the geometric information of each part of the recognition target object 601 in the object data.
- Correct data can be generated from the calculation results of these bounding boxes 703 to 706.
- correct data can also be labeled in units of pixels with respect to the teacher image. For example, by rendering the transmittance of the object image 701 as 0% and the transmittance of the other portions as 100%, rendering the correct label in pixel units for the portion of the object image 701 in the teacher image can do.
- the teacher data generation unit 105 generates a teacher image and correct answer data as described above, and outputs it as teacher data. Thereby, the teacher data generation unit 105 generates the generated teacher image, the coordinate information of the object in the teacher image, the coordinate information of a specific part of the object, the information on the upper category (type) of the object, the object At least one piece of information of the sub-category (attribute) can be output as teacher data.
- the evaluation unit 104 evaluates the identification error characteristic of the classifier 108 that has been learned by the learning unit 102. Thereby, it is possible to automatically specify the upper category and subcategory that the current classifier 108 is not good at identifying.
- the teacher data generation unit 105 is provided in the image processing apparatus 10. As a result, teacher data can be automatically generated for portions where learning is insufficient.
- the teacher data generation unit 105 outputs teacher data including coordinate information of the object in the teacher image.
- this coordinate information it is possible to further increase the teacher data by performing rotation, translation, arbitrary conversion, etc. on the two-dimensional image by using a method of data enlargement (Data ⁇ ⁇ Augmentation).
- the teacher data generation unit 105 generates a teacher image by performing CG synthesis using physical base rendering based on path tracing. Thereby, reflection, refraction, transmission, shadow and the like can be physically reproduced correctly. Therefore, by performing learning of the discriminator 108 using this teacher data, it is possible to realize a discrimination performance that is not significantly different from the performance of learning with an actual image.
- the image processing apparatus 10 uses the evaluation unit 104 and the teacher data generation unit 105 in combination.
- the image processing apparatus specifies and generates teacher data for reducing misrecognition based on the learned discrimination error characteristic of the discriminator 108 and continuously updates the discriminator 108 with the teacher data. it can.
- the output unit 106 provides an end condition for the update learning of the discriminator 108. Thereby, learning can be completed at a suitable timing. As a result, the learning can be continued continuously in a direction to reduce the prediction error without taking a rest for 24 hours 365 days.
- the image processing apparatus 10 evaluates the classifier 108 that identifies an object in the input image and classifies the object into one of a plurality of types.
- the image processing apparatus 10 includes an identification unit 103 and an evaluation unit 104.
- the identifying unit 103 uses the classifier 108 to identify the objects included in each of the plurality of verification images whose types of objects are known, and outputs any of the plurality of types for each verification image.
- the discrimination performance of the discriminator 108 is obtained (step S2003).
- the evaluation unit 104 outputs an evaluation result for the discriminator 108 based on the discrimination performance of the discriminator 108 obtained by the discrimination unit 103 (step S2004). Since it did in this way, the discrimination
- the plurality of types that the classifier 108 identifies and classifies the object includes a plurality of upper categories and a plurality of subcategories obtained by further subdividing each of the plurality of upper categories.
- the identification unit 103 obtains an evaluation value based on a predetermined evaluation criterion for each higher category and subcategory as the identification performance of the classifier 108. Based on this evaluation value, the evaluation unit 104 outputs at least one of a plurality of upper categories and at least one of a plurality of subcategories as an evaluation result for the discriminator 108. Since it did in this way, the evaluation result with respect to the discriminator 108 can be output in an easy-to-understand manner.
- the evaluation unit 104 can also determine a subcategory to be output as an evaluation result for the discriminator 108 based on the evaluation value and the correlation between the subcategories. In this way, since the subcategories having high correlation with each other can be output as the evaluation results, the discrimination performance of the discriminator 108 can be more appropriately evaluated.
- the image processing apparatus 10 further includes a teacher data generation unit 105 and a learning unit 102.
- the teacher data generation unit 105 generates teacher data to be used for learning of the classifier 108 based on the evaluation result for the classifier 108 output from the evaluation unit 104 (step S2005).
- the learning unit 102 learns the discriminator 108 based on the teacher data generated by the teacher data generation unit 105 (step S2002). Since it did in this way, the identification performance of the discrimination device 108 can be improved automatically and reliably.
- the teacher data generation unit 105 synthesizes object data having geometric information and material information and background data having global image or three-dimensional shape information by physical base rendering, thereby using the teacher data used for the teacher data.
- An image is generated (step S5005). Since it did in this way, the teacher image suitable for using for the learning of the discriminator 108 can be reliably generated.
- the teacher data generation unit 105 includes any one of the generated teacher image, the coordinate information of the object in the teacher image, the coordinate information of a specific part of the object, the type information of the object, and the attribute information of the object. Or at least one piece of information can be output as teacher data. By performing learning of the discriminator 108 using this teacher data, the discrimination performance of the discriminator 108 can be reliably improved.
- the teacher data generation unit 105 stores parameter history information 403 used when teacher data is generated in the past. Based on the parameter history information 403, the teacher data generation unit 105 generates teacher data using parameters that are different from previously used parameters. Since it did in this way, the teacher data effective for learning of the discriminator 108 can be produced
- the image processing apparatus 10 further includes an output unit 106.
- the output unit 106 determines whether the learning status of the classifier 108 by the learning unit 102 satisfies a predetermined end condition (S2006).
- the learning unit 102 ends the learning of the classifier 108 when the output unit 106 determines that the learning status of the classifier 108 satisfies the termination condition. Since it did in this way, learning of the discriminator 108 can be complete
- FIG. 9 is a block diagram showing the configuration of the image processing apparatus 80 according to the second embodiment of the present invention.
- the image processing apparatus 80 illustrated in FIG. 9 further includes a notification unit 801 that transmits information to a user 81 who uses the cloud service, and a reception unit 802 that receives information from the user 81.
- a notification unit 801 that transmits information to a user 81 who uses the cloud service
- a reception unit 802 that receives information from the user 81.
- the notification unit 801 notifies the user 81 of the learning status of the classifier 108 by the learning unit 102 based on the information output from the output unit 106.
- the notification from the notification unit 801 includes various information related to learning of the classifier 108 such as the identification error characteristic of the classifier 108 obtained by the evaluation unit 104 and the learning convergence state determined by the output unit 106. included. As a result, the user 81 can grasp the processing status in the image processing apparatus 80.
- FIG. 10 is a diagram illustrating an example of a convergence state of learning of the classifier 108 by the learning unit 102.
- the horizontal axis represents the number of learnings
- the vertical axis represents the prediction error.
- the prediction error decreases each time learning is repeated.
- the prediction error may not be reduced even if the learning is repeated, as indicated by a curve 902.
- Factors include inadequate learning parameters, lack of verification data, lack of data used in teacher data generation, and category design that is difficult to identify in the first place. Is mentioned.
- the notification unit 801 in the present embodiment notifies the user 81 of the factor when the prediction error is not sufficiently reduced as in the curve 902.
- the user 81 adds verification data to the database 107, adds the object data 401 and the background data 402 of FIG. 5 used for generation of teacher data by the teacher data generation unit 105, and the learning unit 102.
- the image processing apparatus 80 is instructed to perform various operations for removing factors such as design changes.
- the receiving unit 802 performs various operations as described above for improving the identification accuracy of an object to be identified by the classifier 108, and optimizes the processing of the image processing apparatus 80. .
- the image processing apparatus 80 is provided with the notification unit 801 that notifies the user of information and the reception unit 802 that receives information from the user.
- the notification unit 801 that notifies the user of information
- the reception unit 802 that receives information from the user.
- the image processing apparatus 80 further includes a notification unit 801 that notifies the user of the learning status of the classifier 108 by the learning unit 102. As a result, the user can check the learning status of the classifier 108.
- the image processing apparatus 80 further includes a receiving unit 802 that receives an instruction from the user and performs an operation for improving the identification accuracy of the object by the classifier 108. Thereby, the user can improve the discriminator 108 as needed.
- FIG. 11 is a block diagram showing a configuration of an image processing system according to the third embodiment of the present invention.
- the image processing system illustrated in FIG. 11 includes an object detection device 1000 that detects an approaching object or an intruder, an image processing device 1001, a camera 1002, and an output device 1003.
- the camera 1002 acquires a video obtained by photographing the periphery of the object detection device 1000 and outputs a video signal based on the video to the object detection device 1000.
- the output device 1003 issues an alarm based on an alarm signal output from the object detection device 1000 using a display or a speaker.
- the object detection apparatus 1000 includes an I / O unit 1004 that functions as an input / output interface that inputs and outputs various data, a CPU 1005 that functions as a processing unit that executes various operations, and a memory 1006.
- the CPU 1005 includes an object detection unit 1008 that performs object detection and a risk determination unit 1009 that determines the risk as functions.
- the object detection apparatus 1000 and the image processing apparatus 1001 are not necessarily installed in the same place.
- the image processing system of this embodiment can be realized by a client server system in which the image processing apparatus 1001 is placed on a server and the object detection apparatus 1000 is operated by a client.
- the processing in the CPU 1005 may be parallel processing using a GPU.
- the image processing apparatus 1001 outputs a discriminator 1007 used in the object detection apparatus 1000 by communicating with the object detection apparatus 1000.
- the image processing apparatus 10 in the first embodiment or the image processing apparatus 80 in the second embodiment can be used. That is, the image processing apparatus 1001 outputs the classifier 108 learned by the learning unit 102 illustrated in FIG. 1 or 9 to the object detection apparatus 1000 as the classifier 1007.
- the discriminator 1007 is stored in the memory 1006 in the object detection apparatus 1000.
- the object detection unit 1008 performs object detection using the discriminator 1007 on the video acquired by the camera 1002. That is, the object detection unit 1008 detects the position and size of the target in the video by identifying the target included in the video input from the camera 1002 with the discriminator 1007.
- object detection in a plurality of frames may be performed using a known tracking method using time-series information.
- the risk level determination unit 1009 determines the risk level of the object detected by the object detection unit 1008 based on a known index such as the degree of approach and the degree of abnormality. As a result, when it is determined that the degree of risk is high, an alarm signal is transmitted to the output device 1003.
- the output device 1003 Upon receiving this warning signal, the output device 1003 issues a warning to the user through a display or a speaker.
- a control signal for performing brake control or steering control may be output to the vehicle instead of outputting an alarm signal to the output device 1003. Good.
- the object detection apparatus 1000 performs object detection using the discriminator 1007 learned by the image processing apparatus 1001. Therefore, an approaching object alarm in the in-vehicle system and an intruder alarm in the monitoring system can be realized. Further, by changing the output signal in accordance with the degree of danger, it is possible to use it not only as an alarm system but also as a control system.
- the discriminator 1007 can be updated online to perform more preferable object detection. For example, it is easy to rewrite the discriminator 1007 in a factory or a store.
- the object detection apparatus 1000 includes an object detection unit 1008 and a risk determination unit 1009.
- the object detection unit 1008 detects an object in the video input from the camera 1002 using the discriminator 1007 that has been learned using the image processing apparatus 1001.
- the risk determination unit 1009 determines the risk level of the object detected by the object detection unit 1008. Since it did in this way, an object can be detected reliably and correctly, and the danger level can be determined.
- the present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the spirit of the present invention.
- the above-described embodiment has been described in detail for easy understanding of the present invention, and is not necessarily limited to the one having all the configurations described.
- a part of the configuration of an embodiment can be replaced with the configuration of another embodiment, and the configuration of another embodiment can be added to the configuration of an embodiment.
- each or all of the above-described configurations, functions, processing units, processing means, and the like may be realized by hardware such as an integrated circuit.
- Information such as programs, data, and files for realizing each configuration and function may be recorded in a recording device such as a memory, a hard disk, or an SSD (Solid State Drive), or an IC card, SD, or the like. You may record on recording media, such as a card
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Multimedia (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Data Mining & Analysis (AREA)
- Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Medical Informatics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- General Engineering & Computer Science (AREA)
- Biodiversity & Conservation Biology (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Image Analysis (AREA)
Abstract
Description
本発明による物体検知装置は、上記の画像処理装置を用いて学習が行われた識別器を用いて、カメラから入力された映像中の物体を検出する物体検出部と、前記物体検出部により検出された物体の危険度を判定する危険度判定部と、を備える。
本発明による画像処理方法は、入力された画像中の対象物を識別して前記対象物を複数の種別のいずれかに分類する識別器を評価するための、コンピュータを用いたものであって、前記コンピュータにより、前記識別器を用いて、前記対象物の種別が既知である複数の検証用画像にそれぞれ含まれる前記対象物を識別して前記複数の種別のいずれかを前記検証用画像ごとに出力することで、前記識別器の識別性能を求め、前記コンピュータにより、求められた前記識別器の識別性能に基づいて、前記識別器に対する評価結果を出力する。
図1は、本発明の第一の実施形態による画像処理装置10の構成を示すブロック図である。図1に示す画像処理装置10は、 入力部101、学習部102、識別部103、評価部104、教師データ生成部105、出力部106を備える。画像処理装置10には、検証データが格納されたデータベース107と、識別器108とが接続されている。なお、データベース107や識別器108は、画像処理装置10の内部にそれぞれ設けられていてもよい。画像処理装置10の各部は、ハードウェアによって構成されてもよいし、コンピュータで実行されるソフトウェアによって構成されていてもよい。また、ハードウェアとソフトウェアを組み合わせたモジュールであってもよい。
次に、本発明の第二の実施形態として、クラウドサービスに適用した実施形態を説明する。
次に、本発明の第三の実施形態として、車載システムにおける接近物警報や、監視システムにおける侵入者警報に適用した実施形態を説明する。
101・・・入力部
102・・・学習部
103・・・識別部
104・・・評価部
105・・・教師データ生成部
106・・・出力部
107・・・データベース
108・・・識別器
401・・・物体データ
402・・・背景データ
403・・・パラメータ履歴情報
404・・・物体設定部
405・・・背景設定部
406・・・パラメータ設定部
407・・・教師画像生成部
408・・・アノテーション部
801・・・通知部
802・・・受信部
1000・・・物体検知装置
Claims (12)
- 入力された画像中の対象物を識別して前記対象物を複数の種別のいずれかに分類する識別器を評価するための画像処理装置であって、
前記識別器を用いて、前記対象物の種別が既知である複数の検証用画像にそれぞれ含まれる前記対象物を識別して前記複数の種別のいずれかを前記検証用画像ごとに出力することで、前記識別器の識別性能を求める識別部と、
前記識別部により求められた前記識別器の識別性能に基づいて、前記識別器に対する評価結果を出力する評価部と、を備える画像処理装置。 - 請求項1に記載の画像処理装置において、
前記複数の種別は、複数の上位カテゴリと、前記複数の上位カテゴリの各々をさらに細分化した複数のサブカテゴリとを含んで構成され、
前記識別部は、前記識別器の識別性能として、所定の評価基準に基づく評価値を前記上位カテゴリおよび前記サブカテゴリごとに求め、
前記評価部は、前記評価値に基づいて、前記複数の上位カテゴリのいずれか少なくとも一つと、前記複数のサブカテゴリのいずれか少なくとも一つとを、前記識別器に対する評価結果として出力する画像処理装置。 - 請求項2に記載の画像処理装置において、
前記評価部は、前記評価値および各サブカテゴリ間の相関性に基づいて、前記識別器に対する評価結果として出力するサブカテゴリを決定する画像処理装置。 - 請求項1乃至3のいずれか一項に記載の画像処理装置において、
前記評価結果に基づいて、前記識別器の学習に用いるための教師データを生成する教師データ生成部と、
前記教師データに基づいて、前記識別器の学習を行う学習部と、をさらに備える画像処理装置。 - 請求項4に記載の画像処理装置において、
前記教師データ生成部は、幾何情報およびマテリアル情報を有する物体データと、全球画像または三次元形状情報を有する背景データとを、物理ベースレンダリングで合成することにより、前記教師データに用いられる教師画像を生成する画像処理装置。 - 請求項4に記載の画像処理装置において、
前記教師データ生成部は、教師画像と、前記教師画像における前記対象物の座標情報、前記対象物のうち特定の部位の座標情報、前記対象物の種別情報、前記対象物の属性情報のいずれか少なくとも一つの情報とを、前記教師データとして出力する画像処理装置。 - 請求項4に記載の画像処理装置において、
前記教師データ生成部は、過去に前記教師データを生成した際に用いたパラメータの履歴情報を記憶しており、前記履歴情報に基づいて、過去に使用済みのパラメータとは異なるパラメータを用いて前記教師データを生成する画像処理装置。 - 請求項4に記載の画像処理装置において、
前記学習部による前記識別器の学習状況が所定の終了条件を満たしたか否かを判定する出力部をさらに備え、
前記学習部は、前記出力部により前記識別器の学習状況が前記終了条件を満たしたと判定された場合に、前記識別器の学習を終了する画像処理装置。 - 請求項4に記載の画像処理装置において、
前記学習部による前記識別器の学習状況をユーザに通知する通知部をさらに備える画像処理装置。 - 請求項9に記載の画像処理装置において、
前記ユーザからの指示を受けて、前記識別器による前記対象物の識別精度を向上させるための動作を行う受信部をさらに備える画像処理装置。 - 請求項4に記載の画像処理装置を用いて学習が行われた識別器を用いて、カメラから入力された映像中の物体を検出する物体検出部と、
前記物体検出部により検出された物体の危険度を判定する危険度判定部と、を備える物体検知装置。 - 入力された画像中の対象物を識別して前記対象物を複数の種別のいずれかに分類する識別器を評価するための、コンピュータを用いた画像処理方法であって、
前記コンピュータにより、前記識別器を用いて、前記対象物の種別が既知である複数の検証用画像にそれぞれ含まれる前記対象物を識別して前記複数の種別のいずれかを前記検証用画像ごとに出力することで、前記識別器の識別性能を求め、
前記コンピュータにより、求められた前記識別器の識別性能に基づいて、前記識別器に対する評価結果を出力する画像処理方法。
Priority Applications (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2017509108A JP6546271B2 (ja) | 2015-04-02 | 2015-04-02 | 画像処理装置、物体検知装置、画像処理方法 |
US15/562,535 US10373024B2 (en) | 2015-04-02 | 2015-04-02 | Image processing device, object detection device, image processing method |
PCT/JP2015/060482 WO2016157499A1 (ja) | 2015-04-02 | 2015-04-02 | 画像処理装置、物体検知装置、画像処理方法 |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/JP2015/060482 WO2016157499A1 (ja) | 2015-04-02 | 2015-04-02 | 画像処理装置、物体検知装置、画像処理方法 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2016157499A1 true WO2016157499A1 (ja) | 2016-10-06 |
Family
ID=57004016
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2015/060482 WO2016157499A1 (ja) | 2015-04-02 | 2015-04-02 | 画像処理装置、物体検知装置、画像処理方法 |
Country Status (3)
Country | Link |
---|---|
US (1) | US10373024B2 (ja) |
JP (1) | JP6546271B2 (ja) |
WO (1) | WO2016157499A1 (ja) |
Cited By (29)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2018081549A (ja) * | 2016-11-17 | 2018-05-24 | 清水建設株式会社 | モデル作成装置、端末装置、危険箇所発見システム及びプログラム |
WO2018135605A1 (ja) * | 2017-01-23 | 2018-07-26 | パナソニックIpマネジメント株式会社 | イベント予測システム、イベント予測方法、プログラム、及び移動体 |
WO2018135509A1 (ja) * | 2017-01-23 | 2018-07-26 | パナソニックIpマネジメント株式会社 | イベント予測システム、イベント予測方法、プログラムおよびそれを記録した記録媒体 |
KR101886754B1 (ko) * | 2017-05-04 | 2018-09-10 | 국방과학연구소 | 머신 러닝을 위한 학습 이미지 생성 장치 및 방법 |
JP2018146587A (ja) * | 2017-03-08 | 2018-09-20 | 清華大学Tsinghua University | 検査機器と車両の画像を分割する方法 |
JP2018152106A (ja) * | 2018-05-02 | 2018-09-27 | 株式会社日立国際電気 | 水上侵入検知システムおよびその方法 |
JP2018163554A (ja) * | 2017-03-27 | 2018-10-18 | 富士通株式会社 | 画像処理装置、画像処理方法、画像処理プログラム、及び教師データ生成方法 |
JP2018169746A (ja) * | 2017-03-29 | 2018-11-01 | 本田技研工業株式会社 | 物体認識装置および物体認識方法 |
JP2018169672A (ja) * | 2017-03-29 | 2018-11-01 | 三菱電機インフォメーションシステムズ株式会社 | 教師画像を生成する方法、コンピュータおよびプログラム |
WO2018224874A1 (en) | 2017-06-06 | 2018-12-13 | PlusAI Corp | Method and system for distributed learning and adaptation in autonomous driving vehicles |
JP2019020820A (ja) * | 2017-07-12 | 2019-02-07 | 株式会社日立製作所 | 映像認識システム |
KR101947650B1 (ko) * | 2017-11-14 | 2019-05-20 | 국방과학연구소 | 게임 엔진 기반의 머신 러닝 학습이미지 생성 장치 및 방법 |
KR20190054782A (ko) * | 2017-11-14 | 2019-05-22 | 국방과학연구소 | 머신 러닝의 학습성능 강화장치 및 방법 |
WO2019150918A1 (ja) * | 2018-02-02 | 2019-08-08 | ソニー株式会社 | 情報処理装置、情報処理方法、プログラム、及び移動体 |
CN110503126A (zh) * | 2018-05-18 | 2019-11-26 | 罗伯特·博世有限公司 | 用于改善分类器的训练的方法和设备 |
JP2020013467A (ja) * | 2018-07-20 | 2020-01-23 | 株式会社壽屋 | 分類器生成方法、機械学習用画像及び学習データ生成方法 |
JP2020067669A (ja) * | 2018-10-19 | 2020-04-30 | 株式会社ファブリカコミュニケーションズ | 情報処理装置及びプログラム |
JP2020091813A (ja) * | 2018-12-07 | 2020-06-11 | 公立大学法人会津大学 | ニューラルネットワークの学習方法、コンピュータプログラム及びコンピュータ装置 |
WO2020050934A3 (en) * | 2018-08-10 | 2020-06-18 | Nec Laboratories America, Inc. | Parametric top-view representation of scenes |
JP2020107254A (ja) * | 2018-12-28 | 2020-07-09 | 株式会社AI−feed | 画像検査装置、学習済みモデル生成装置、画像検査システム、画像検査用プログラム、学習済みモデル生成用プログラム、および学習済みモデル |
JP2020166473A (ja) * | 2019-03-29 | 2020-10-08 | 三井情報株式会社 | 学習済みモデル生成装置、プログラム及び学習済みモデル生成システム |
JP6775697B1 (ja) * | 2019-06-20 | 2020-10-28 | 三菱電機株式会社 | 学習用データ生成装置、学習用データ生成方法、学習用データ生成プログラム、学習装置、学習方法、学習プログラム、推論装置、推論方法、推論プログラム、学習システム、及び推論システム |
KR20200133555A (ko) * | 2019-05-20 | 2020-11-30 | 넷마블 주식회사 | 기계 학습을 위한 데이터 수집 방법 |
JP2020201746A (ja) * | 2019-06-11 | 2020-12-17 | トヨタ自動車株式会社 | 距離推定装置、距離推定方法及び距離推定用コンピュータプログラム |
JP2021058232A (ja) * | 2019-10-02 | 2021-04-15 | コニカミノルタ株式会社 | 超音波画像診断用訓練装置、超音波画像診断装置、識別モデルの訓練方法、プログラム及び超音波診断装置 |
JPWO2021095101A1 (ja) * | 2019-11-11 | 2021-05-20 | ||
JPWO2021166157A1 (ja) * | 2020-02-20 | 2021-08-26 | ||
WO2022186110A1 (ja) * | 2021-03-01 | 2022-09-09 | 富士フイルム株式会社 | 機械学習システム、認識器、学習方法、及びプログラム |
US12039445B2 (en) | 2022-12-07 | 2024-07-16 | Plusai, Inc. | Method and system for on-the-fly object labeling via cross modality validation in autonomous driving vehicles |
Families Citing this family (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11042155B2 (en) | 2017-06-06 | 2021-06-22 | Plusai Limited | Method and system for closed loop perception in autonomous driving vehicles |
US11392133B2 (en) * | 2017-06-06 | 2022-07-19 | Plusai, Inc. | Method and system for object centric stereo in autonomous driving vehicles |
JP7345236B2 (ja) * | 2017-11-10 | 2023-09-15 | 株式会社小松製作所 | 作業車両の動作を推定するための方法、システム、学習済みの分類モデルの製造方法、学習データ、及び学習データの製造方法 |
US10304208B1 (en) | 2018-02-12 | 2019-05-28 | Avodah Labs, Inc. | Automated gesture identification using neural networks |
WO2019157344A1 (en) | 2018-02-12 | 2019-08-15 | Avodah Labs, Inc. | Real-time gesture recognition method and apparatus |
US10289903B1 (en) * | 2018-02-12 | 2019-05-14 | Avodah Labs, Inc. | Visual sign language translation training device and method |
US11417109B1 (en) * | 2018-03-20 | 2022-08-16 | Amazon Technologies, Inc. | Network-based vehicle event detection system |
CN108921161B (zh) * | 2018-06-08 | 2021-03-23 | Oppo广东移动通信有限公司 | 模型训练方法、装置、电子设备和计算机可读存储介质 |
JP6697042B2 (ja) * | 2018-08-31 | 2020-05-20 | ソニーセミコンダクタソリューションズ株式会社 | 固体撮像システム、固体撮像方法及びプログラム |
CN111160411B (zh) * | 2019-12-11 | 2023-09-29 | 东软集团股份有限公司 | 分类模型训练方法、图像处理方法、装置、介质及设备 |
CN111311578B (zh) * | 2020-02-17 | 2024-05-03 | 腾讯科技(深圳)有限公司 | 基于人工智能的对象分类方法以及装置、医学影像设备 |
CN114549968A (zh) * | 2022-01-11 | 2022-05-27 | 深圳绿米联创科技有限公司 | 目标检测方法、装置以及电子设备 |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2002259974A (ja) * | 2001-03-02 | 2002-09-13 | Denso Corp | 画像領域抽出装置、プログラム及び記録媒体 |
JP2009282824A (ja) * | 2008-05-23 | 2009-12-03 | Toyota Central R&D Labs Inc | 感情推定装置及びプログラム |
JP2012123642A (ja) * | 2010-12-08 | 2012-06-28 | Toyota Central R&D Labs Inc | 画像識別装置及びプログラム |
JP2014115920A (ja) * | 2012-12-12 | 2014-06-26 | Casio Comput Co Ltd | 多クラス識別器、方法、及びプログラム |
JP2014178229A (ja) * | 2013-03-15 | 2014-09-25 | Dainippon Screen Mfg Co Ltd | 教師データ作成方法、画像分類方法および画像分類装置 |
Family Cites Families (25)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6262742B1 (en) * | 1999-03-03 | 2001-07-17 | Discreet Logic Inc. | Generating image data |
US6940545B1 (en) * | 2000-02-28 | 2005-09-06 | Eastman Kodak Company | Face detecting camera and method |
US9955551B2 (en) * | 2002-07-12 | 2018-04-24 | Yechezkal Evan Spero | Detector controlled illuminating system |
US7508961B2 (en) * | 2003-03-12 | 2009-03-24 | Eastman Kodak Company | Method and system for face detection in digital images |
US8593542B2 (en) * | 2005-12-27 | 2013-11-26 | DigitalOptics Corporation Europe Limited | Foreground/background separation using reference images |
US7574016B2 (en) * | 2003-06-26 | 2009-08-11 | Fotonation Vision Limited | Digital image processing using face detection information |
US7551755B1 (en) * | 2004-01-22 | 2009-06-23 | Fotonation Vision Limited | Classification and organization of consumer digital images using workflow, and face detection and recognition |
US7868904B2 (en) * | 2005-04-01 | 2011-01-11 | Canon Kabushiki Kaisha | Image processing method and image processing apparatus |
US8446494B2 (en) * | 2008-02-01 | 2013-05-21 | Hewlett-Packard Development Company, L.P. | Automatic redeye detection based on redeye and facial metric values |
WO2009117607A1 (en) * | 2008-03-19 | 2009-09-24 | The Trustees Of Columbia University In The City Of New York | Methods, systems, and media for automatically classifying face images |
GB2459251A (en) * | 2008-04-01 | 2009-10-21 | Sharp Kk | Semiconductor nanowire devices |
JP5361530B2 (ja) * | 2009-05-20 | 2013-12-04 | キヤノン株式会社 | 画像認識装置、撮像装置及び画像認識方法 |
US8805707B2 (en) * | 2009-12-31 | 2014-08-12 | Hartford Fire Insurance Company | Systems and methods for providing a safety score associated with a user location |
US8751215B2 (en) * | 2010-06-04 | 2014-06-10 | Microsoft Corporation | Machine based sign language interpreter |
US9916538B2 (en) * | 2012-09-15 | 2018-03-13 | Z Advanced Computing, Inc. | Method and system for feature detection |
US9153084B2 (en) * | 2012-03-14 | 2015-10-06 | Flextronics Ap, Llc | Destination and travel information application |
US8817120B2 (en) * | 2012-05-31 | 2014-08-26 | Apple Inc. | Systems and methods for collecting fixed pattern noise statistics of image data |
US8953882B2 (en) * | 2012-05-31 | 2015-02-10 | Apple Inc. | Systems and methods for determining noise statistics of image data |
US9031317B2 (en) * | 2012-09-18 | 2015-05-12 | Seiko Epson Corporation | Method and apparatus for improved training of object detecting system |
US9672653B2 (en) * | 2012-12-14 | 2017-06-06 | Nvidia Corporation | Stereo viewpoint graphics processing subsystem and method of sharing geometry data between stereo images in screen-spaced processing |
JP6188345B2 (ja) | 2013-02-15 | 2017-08-30 | キヤノン株式会社 | 情報処理装置、情報処理方法 |
US9269012B2 (en) * | 2013-08-22 | 2016-02-23 | Amazon Technologies, Inc. | Multi-tracker object tracking |
US9443169B2 (en) * | 2014-02-21 | 2016-09-13 | Xerox Corporation | Object classification with constrained multiple instance support vector machine |
US9595130B2 (en) * | 2014-06-17 | 2017-03-14 | Chief Architect Inc. | Virtual model navigation methods and apparatus |
US10599926B2 (en) * | 2015-12-16 | 2020-03-24 | Regents Of The University Of Minnesota | Automated detection of nitrogen deficiency in crop |
-
2015
- 2015-04-02 WO PCT/JP2015/060482 patent/WO2016157499A1/ja active Application Filing
- 2015-04-02 US US15/562,535 patent/US10373024B2/en not_active Expired - Fee Related
- 2015-04-02 JP JP2017509108A patent/JP6546271B2/ja not_active Expired - Fee Related
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2002259974A (ja) * | 2001-03-02 | 2002-09-13 | Denso Corp | 画像領域抽出装置、プログラム及び記録媒体 |
JP2009282824A (ja) * | 2008-05-23 | 2009-12-03 | Toyota Central R&D Labs Inc | 感情推定装置及びプログラム |
JP2012123642A (ja) * | 2010-12-08 | 2012-06-28 | Toyota Central R&D Labs Inc | 画像識別装置及びプログラム |
JP2014115920A (ja) * | 2012-12-12 | 2014-06-26 | Casio Comput Co Ltd | 多クラス識別器、方法、及びプログラム |
JP2014178229A (ja) * | 2013-03-15 | 2014-09-25 | Dainippon Screen Mfg Co Ltd | 教師データ作成方法、画像分類方法および画像分類装置 |
Non-Patent Citations (1)
Title |
---|
NAOYA KOKETSU ET AL.: "Improvement in Human Detection by Generation Addition Training of Human Silhouette", IPSJ SIG NOTES, HEISEI 22 NENDO 3 [ CD-ROM, 15 October 2010 (2010-10-15) * |
Cited By (43)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2018081549A (ja) * | 2016-11-17 | 2018-05-24 | 清水建設株式会社 | モデル作成装置、端末装置、危険箇所発見システム及びプログラム |
WO2018135605A1 (ja) * | 2017-01-23 | 2018-07-26 | パナソニックIpマネジメント株式会社 | イベント予測システム、イベント予測方法、プログラム、及び移動体 |
WO2018135509A1 (ja) * | 2017-01-23 | 2018-07-26 | パナソニックIpマネジメント株式会社 | イベント予測システム、イベント予測方法、プログラムおよびそれを記録した記録媒体 |
JP2018146587A (ja) * | 2017-03-08 | 2018-09-20 | 清華大学Tsinghua University | 検査機器と車両の画像を分割する方法 |
JP7011146B2 (ja) | 2017-03-27 | 2022-01-26 | 富士通株式会社 | 画像処理装置、画像処理方法、画像処理プログラム、及び教師データ生成方法 |
JP2018163554A (ja) * | 2017-03-27 | 2018-10-18 | 富士通株式会社 | 画像処理装置、画像処理方法、画像処理プログラム、及び教師データ生成方法 |
JP2018169746A (ja) * | 2017-03-29 | 2018-11-01 | 本田技研工業株式会社 | 物体認識装置および物体認識方法 |
JP2018169672A (ja) * | 2017-03-29 | 2018-11-01 | 三菱電機インフォメーションシステムズ株式会社 | 教師画像を生成する方法、コンピュータおよびプログラム |
KR101886754B1 (ko) * | 2017-05-04 | 2018-09-10 | 국방과학연구소 | 머신 러닝을 위한 학습 이미지 생성 장치 및 방법 |
EP3635497A4 (en) * | 2017-06-06 | 2021-04-21 | PlusAI Corp | LEARNING AND ADAPTATION PROCESS AND SYSTEM DISTRIBUTED IN AUTONOMOUS DRIVING VEHICLES |
WO2018224874A1 (en) | 2017-06-06 | 2018-12-13 | PlusAI Corp | Method and system for distributed learning and adaptation in autonomous driving vehicles |
JP2019020820A (ja) * | 2017-07-12 | 2019-02-07 | 株式会社日立製作所 | 映像認識システム |
KR20190054782A (ko) * | 2017-11-14 | 2019-05-22 | 국방과학연구소 | 머신 러닝의 학습성능 강화장치 및 방법 |
KR102053202B1 (ko) | 2017-11-14 | 2019-12-06 | 국방과학연구소 | 머신 러닝의 학습성능 강화장치 및 방법 |
KR101947650B1 (ko) * | 2017-11-14 | 2019-05-20 | 국방과학연구소 | 게임 엔진 기반의 머신 러닝 학습이미지 생성 장치 및 방법 |
WO2019150918A1 (ja) * | 2018-02-02 | 2019-08-08 | ソニー株式会社 | 情報処理装置、情報処理方法、プログラム、及び移動体 |
US11615628B2 (en) | 2018-02-02 | 2023-03-28 | Sony Corporation | Information processing apparatus, information processing method, and mobile object |
JP2018152106A (ja) * | 2018-05-02 | 2018-09-27 | 株式会社日立国際電気 | 水上侵入検知システムおよびその方法 |
CN110503126A (zh) * | 2018-05-18 | 2019-11-26 | 罗伯特·博世有限公司 | 用于改善分类器的训练的方法和设备 |
JP2020013467A (ja) * | 2018-07-20 | 2020-01-23 | 株式会社壽屋 | 分類器生成方法、機械学習用画像及び学習データ生成方法 |
WO2020050934A3 (en) * | 2018-08-10 | 2020-06-18 | Nec Laboratories America, Inc. | Parametric top-view representation of scenes |
JP2020067669A (ja) * | 2018-10-19 | 2020-04-30 | 株式会社ファブリカコミュニケーションズ | 情報処理装置及びプログラム |
JP2020091813A (ja) * | 2018-12-07 | 2020-06-11 | 公立大学法人会津大学 | ニューラルネットワークの学習方法、コンピュータプログラム及びコンピュータ装置 |
JP7230324B2 (ja) | 2018-12-07 | 2023-03-01 | 公立大学法人会津大学 | ニューラルネットワークの学習方法、コンピュータプログラム及びコンピュータ装置 |
JP2020107254A (ja) * | 2018-12-28 | 2020-07-09 | 株式会社AI−feed | 画像検査装置、学習済みモデル生成装置、画像検査システム、画像検査用プログラム、学習済みモデル生成用プログラム、および学習済みモデル |
JP7340806B2 (ja) | 2018-12-28 | 2023-09-08 | 株式会社Ridge-i | 画像検査装置、画像検査システム、および画像検査用プログラム |
JP2020166473A (ja) * | 2019-03-29 | 2020-10-08 | 三井情報株式会社 | 学習済みモデル生成装置、プログラム及び学習済みモデル生成システム |
JP7267068B2 (ja) | 2019-03-29 | 2023-05-01 | 三井情報株式会社 | 学習済みモデル生成装置、プログラム及び学習済みモデル生成システム |
KR20200133555A (ko) * | 2019-05-20 | 2020-11-30 | 넷마블 주식회사 | 기계 학습을 위한 데이터 수집 방법 |
KR102199467B1 (ko) * | 2019-05-20 | 2021-01-07 | 넷마블 주식회사 | 기계 학습을 위한 데이터 수집 방법 |
JP2020201746A (ja) * | 2019-06-11 | 2020-12-17 | トヨタ自動車株式会社 | 距離推定装置、距離推定方法及び距離推定用コンピュータプログラム |
JP7003972B2 (ja) | 2019-06-11 | 2022-01-21 | トヨタ自動車株式会社 | 距離推定装置、距離推定方法及び距離推定用コンピュータプログラム |
JP6775697B1 (ja) * | 2019-06-20 | 2020-10-28 | 三菱電機株式会社 | 学習用データ生成装置、学習用データ生成方法、学習用データ生成プログラム、学習装置、学習方法、学習プログラム、推論装置、推論方法、推論プログラム、学習システム、及び推論システム |
JP2021058232A (ja) * | 2019-10-02 | 2021-04-15 | コニカミノルタ株式会社 | 超音波画像診断用訓練装置、超音波画像診断装置、識別モデルの訓練方法、プログラム及び超音波診断装置 |
JP7427902B2 (ja) | 2019-10-02 | 2024-02-06 | コニカミノルタ株式会社 | 超音波画像診断用訓練装置、超音波画像診断装置、識別モデルの訓練方法、プログラム及び超音波診断装置 |
JPWO2021095101A1 (ja) * | 2019-11-11 | 2021-05-20 | ||
WO2021095101A1 (ja) * | 2019-11-11 | 2021-05-20 | 日本電信電話株式会社 | 学習装置、検知装置、学習方法、及び異常検知方法 |
JP7338698B2 (ja) | 2019-11-11 | 2023-09-05 | 日本電信電話株式会社 | 学習装置、検知装置、学習方法、及び異常検知方法 |
WO2021166157A1 (ja) * | 2020-02-20 | 2021-08-26 | 三菱電機株式会社 | 物体認識システム検証装置 |
JPWO2021166157A1 (ja) * | 2020-02-20 | 2021-08-26 | ||
JP7199592B2 (ja) | 2020-02-20 | 2023-01-05 | 三菱電機株式会社 | 物体認識システム検証装置 |
WO2022186110A1 (ja) * | 2021-03-01 | 2022-09-09 | 富士フイルム株式会社 | 機械学習システム、認識器、学習方法、及びプログラム |
US12039445B2 (en) | 2022-12-07 | 2024-07-16 | Plusai, Inc. | Method and system for on-the-fly object labeling via cross modality validation in autonomous driving vehicles |
Also Published As
Publication number | Publication date |
---|---|
US10373024B2 (en) | 2019-08-06 |
JPWO2016157499A1 (ja) | 2017-12-07 |
JP6546271B2 (ja) | 2019-07-17 |
US20180107901A1 (en) | 2018-04-19 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2016157499A1 (ja) | 画像処理装置、物体検知装置、画像処理方法 | |
JP6873237B2 (ja) | 画像ベースの車両損害評価方法、装置、およびシステム、ならびに電子デバイス | |
US20210081698A1 (en) | Systems and methods for physical object analysis | |
KR102418446B1 (ko) | 픽쳐 기반의 차량 손해 평가 방법 및 장치, 및 전자 디바이스 | |
KR102138082B1 (ko) | 다수의 이미지 일치성을 바탕으로 보험클레임 사기 방지를 실현하는 방법, 시스템, 기기 및 판독 가능 저장매체 | |
CN108388834A (zh) | 利用循环神经网络和级联特征映射的对象检测 | |
CN112560774A (zh) | 一种障碍物位置检测方法、装置、设备和存储介质 | |
US11113582B2 (en) | Method and system for facilitating detection and identification of vehicle parts | |
JP2008130086A (ja) | 分類境界確定方法及び分類境界確定装置 | |
JP2021534494A (ja) | 自律型車両用のカメラ評価技術 | |
US11250279B2 (en) | Generative adversarial network models for small roadway object detection | |
Toprak et al. | Conditional weighted ensemble of transferred models for camera based onboard pedestrian detection in railway driver support systems | |
JP7165353B2 (ja) | 画像特徴量出力装置、画像認識装置、画像特徴量出力プログラム、及び画像認識プログラム | |
EP4341913A2 (en) | System for detection and management of uncertainty in perception systems, for new object detection and for situation anticipation | |
Calitz et al. | Automated license plate recognition using existing university infrastructure and different camera angles | |
CN117786520B (zh) | 目标感知模型的训练方法和应用、无人车和存储介质 | |
US11580723B2 (en) | Scene-aware object detection | |
US11514530B2 (en) | Image processing system using convolutional neural networks | |
Baek et al. | Fast and reliable two-wheeler detection algorithm for blind spot detection systems | |
Calitz et al. | The African Journal of Information Systems | |
Sestakovskis | Computer Vision-Based Dangerous Riding Behaviors Detection | |
JP2021152826A (ja) | 情報処理装置、被写体分類方法、及び被写体分類プログラム | |
KR20240015836A (ko) | 중첩된 객체를 분리하기 위한 방법 및 장치 | |
Datondji | Wide-baseline Omnidirectional Stereovision for Intersection Monitoring | |
CN113392735A (zh) | 示警方法及相关设备 |
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: 15887645 Country of ref document: EP Kind code of ref document: A1 |
|
ENP | Entry into the national phase |
Ref document number: 2017509108 Country of ref document: JP Kind code of ref document: A |
|
WWE | Wipo information: entry into national phase |
Ref document number: 15562535 Country of ref document: US |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 15887645 Country of ref document: EP Kind code of ref document: A1 |