Information processing device, information processing method, and program

Abstract

To provide an information processing device and a method and the like capable of setting the range of object expansion or randomization when generating training data used for machine learning.SOLUTION: An information processing device 10 includes: an arrangement unit 11 that arranges multiple objects in an n-dimensional state space with n attributes; a grouping unit 12 that groups the objects according to a desired task; an area determination unit 13 that determines the region of each group that marks the boundary between groups in state space; and a range determination unit 14 that determines the area on the obtained state space as the range of expansion or randomization for each group.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to an information processing device, an information processing method, and a program, and more particularly to an information processing device, an information processing method, and a program for generating learning data for training a recognizer that recognizes a target object. [Background technology]

[0002] A large amount of training data is required to train a recognizer to recognize target objects. Furthermore, data diversity is also required for robust recognition. To generate large amounts of diverse training data, techniques such as data augmentation and domain randomization are used.

[0003] Patent Document 1 discloses a machine learning device that preferentially selects a predetermined number of images with low classification reliability and / or high classification reliability from among images used in past machine learning as images to be used for machine learning, randomly selects a predetermined number of these selected images as training data, performs new machine learning using this training data, and uses images from the selected images other than those selected as training data as test data, and calculates the image classification accuracy using machine learning features calculated using this test data and training data related to the test data. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2017-224184 Summary of the Invention [Problem to be solved by the invention]

[0005] When expanding or randomizing training data, the range of object expansion or randomization is not defined. This range is not defined for each attribute (e.g., color, size), for example. Therefore, if training data is expanded or randomized beyond the allowable range, the recognition rate of a recognizer created by training based on that data will decrease. If the data is too diverse, the number of misrecognitions (e.g., false positives) will increase.

[0006] The present disclosure has been made to solve such problems, and aims to provide an information processing device, method, etc. that can set the range of object expansion or randomization when generating training data to be used in machine learning. [Means for solving the problem]

[0007] An information processing device according to a first aspect of the present disclosure includes: an arrangement unit that arranges a plurality of objects in an n-dimensional state space based on n attributes; a grouping unit for grouping objects according to a target task; an area determination unit that determines an area of ​​each group that indicates a boundary between the groups in the state space; and a range determination unit that determines the determined region in the state space as the range for expanding or randomizing each group.

[0008] An information processing method according to a second aspect of the present disclosure includes: Multiple objects are placed in an n-dimensional state space with n attributes, Group objects according to the target task, determining a region of each group that indicates a boundary between the groups in the state space; The region in the state space thus obtained is determined as the range for expansion or randomization of each group.

[0009] A program according to a third aspect of the present disclosure includes: Multiple objects are placed in an n-dimensional state space with n attributes, Group objects according to the target task, determining a region of each group that indicates a boundary between the groups in the state space; The computer determines the region in the state space thus obtained as the range for expansion or randomization of each group. [Effects of the Invention]

[0010] According to the present disclosure, it is possible to provide an information processing device, method, etc. that can set the range of object expansion or randomization when generating learning data used in machine learning. [Brief explanation of the drawings]

[0011] [Figure 1] 1 is a block diagram showing a configuration example of an information processing device according to a first embodiment; [Figure 2] 3 is a flowchart showing an information processing method according to the first embodiment. [Figure 3] 10 is a flowchart showing a method for generating training data according to the second embodiment. [Figure 4] FIG. 10 is a diagram illustrating a task 1 according to the first embodiment. [Figure 5] FIG. 10 is a diagram illustrating an example of learning data generated in the case of task 1 according to the first embodiment. [Figure 6] FIG. 10 is a diagram illustrating a task 2 according to the second embodiment. [Figure 7] FIG. 10 is a diagram illustrating an example of learning data generated in the case of task 2 according to the second embodiment. [Figure 8] FIG. 1 illustrates an example of a hardware configuration of an information processing device according to an embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0012] Embodiment 1 Hereinafter, an embodiment of the present invention will be described with reference to the drawings. The present disclosure relates to an information processing device (also referred to as an expansion or randomization range setting device) that can automatically set a range (parameter) for expanding or randomizing training data used in machine learning. The present disclosure also relates to a training data generation device that generates training data within the automatically set expansion or randomization range.

[0013] A typical method for generating training data using a simulator is as follows. (1) Scan the object and create a model. (2) The model is randomly placed in the simulator, and training data with annotations (bounding box, position, orientation, etc.) is generated. (3) Train the recognizer using the generated training data to recognize real-world objects.

[0014] Robust object recognition requires diversity in training data. For this reason, techniques such as data augmentation and domain randomization are used. Data augmentation refers to increasing the amount of data by modifying existing data or generating artificial data from existing data. Examples of data augmentation include changing the contrast or size of images. Domain randomization is a method for addressing the gap between simulation and reality, generating diverse simulation data to incorporate data closer to reality. Examples of domain randomization include changing the position, orientation, and type of object, or changing lighting conditions. Other methods that have been proposed include generating training data by copying and pasting photos.

[0015] By using the method described above to generate diverse data and perform machine learning, it is possible to create a recognizer that can robustly recognize objects. However, if the data is too diverse, a recognizer designed to recognize a specific object may mistakenly recognize a different object as the specific object, resulting in false positives. Furthermore, when generating training data that includes a target object to be recognized and background objects that should not be recognized, the diverse training data may include training data in which a background object similar to the target object is placed near the target object, resulting in a lower recognition rate.

[0016] As described above, up until now, when expanding or randomizing training data, parameters such as variance and threshold have not been set. Therefore, there is a problem in that it is unclear to what extent data can be processed or to what extent the simulation environment can be diversified. Therefore, this disclosure proposes setting parameters for expanding or randomizing training data according to the content of the object recognition task.

[0017] FIG. 1 is a block diagram illustrating a configuration example of an information processing device according to a first embodiment. The information processing device 10 can be used to set parameters for expanding or randomizing training data. The information processing device 10 can also be used to generate a large amount of training data for recognizing target objects. The information processing device 10 includes a memory, a processor, and the like. The information processing device 10 can also include a GPU (Graphics Processing Unit). As shown in FIG. 1 , the information processing device 10 includes an arrangement unit 11, a grouping unit 12, a region determination unit 13, a range determination unit 14, and a generation unit 15.

[0018] The arrangement unit 11 arranges multiple objects in an n-dimensional state space with n attributes, where n is an arbitrary integer. For example, a state space model such as time series analysis may be used. The grouping unit 12, as will be described in detail later, groups one or more objects according to a target task. The region determination unit 13 determines a region for each group that indicates a boundary between the groups in the state space. The range determination unit 14 determines the determined region in the state space as a range for expansion or randomization of each group. In some embodiments, the information processing device 10 may further include a generation unit that expands or randomizes objects within the determined range to generate training data for object recognition. In other embodiments, the information processing device 10 may further include a learning unit that performs machine learning using the generated training data for object recognition to generate an object recognition model. For example, machine learning such as a support vector machine or a neural network may be used.

[0019] FIG. 2 is a flowchart illustrating the information processing method according to the first embodiment. The arrangement unit 11 of the information processing device 10 arranges a plurality of objects in an n-dimensional state with n attributes (step S101). The objects are arranged in a simulator. Here, the attributes of an object refer to the characteristics or properties of the object. Examples of attributes include, but are not limited to, color, length, texture, material, pattern, etc., or combinations thereof. In some embodiments, for example, a product name, trademark, or manufacturer name may be used as an attribute.

[0020] Next, the grouping unit 12 of the information processing device 10 groups the objects according to the target task (step S102). The target task indicates how multiple objects are to be classified. Examples of target tasks include, but are not limited to, a task to classify multiple objects by product type (e.g., "Task 1" described below) or a task to classify by size (e.g., "Task 2" described below). Examples of groups to be classified include, but are not limited to, a case where there are three target object groups, or a case where there is a target object group and a background object group. The areas of each group may or may not be adjacent to each other at a boundary. Note that a target object is an object to be recognized in an image, and a background object is an object in an image that does not need to be recognized.

[0021] The information processing device 100 determines the area of ​​each group (area between groups) in the state space (step S103). The area in the state space determined in step S103 is set as the range of expansion or randomization for each group (step S104). This makes it possible to automatically set parameters such as variance and threshold for expansion or randomization, and also to recognize how much data should be processed or how much the simulation environment should be diversified (for example, whether the brightness of lighting should be changed).

[0022] The objects are randomized within the randomization range set in step S104, and training data is generated (step S105). A recognizer created by machine learning using the training data generated in this way suppresses the above-mentioned misrecognition and improves the recognition rate.

[0023] Embodiment 2 Next, a method for generating training data according to the second embodiment will be described with reference to FIG. A plurality of objects are arranged in an n-dimensional state space with n attributes (step S201). The plurality of objects are grouped according to the target task (step S202). The distribution of similarities between each target object within the group is calculated (step S203). In addition, the distribution of dissimilarity for each attribute between one or more target objects within the group and one or more objects outside the group is calculated (step S204).

[0024] Here, "similarity" refers to the degree or distribution of similarity between one or more target objects in a group. For example, the degree of similarity can be expressed as a numerical value between 0 and 1. In this case, the more similar a recognition target object is to other recognition target objects in the group, the closer to 1 the similarity value will be, and conversely, the less similar a recognition target object is to other recognition target objects in the group, the closer to 0 the similarity value will be. Furthermore, any number between 0 and 1 may be set as the threshold. Furthermore, in some embodiments, the similarity may be expressed as a distribution.

[0025] "Dissimilarity" refers to the degree or distribution of difference between a recognition target object within a group and an object outside the group. For example, the dissimilarity can be expressed as a numerical value between 0 and 1. In this case, the more different (dissimilar) the recognition target object is from other objects outside the group, the closer the dissimilarity value to 1, for example, and conversely, the more similar (dissimilar) the recognition target object is to other objects outside the group, the closer the dissimilarity value to 0. Furthermore, any number between 0 and 1 may be set as the threshold. In some embodiments, Dissimilarity can be expressed as a distribution.

[0026] Furthermore, if a group contains three or more objects, correlations between the objects may be calculated as necessary. To avoid misrecognizing unknown objects, objects that are dissimilar (distant) from the target object may also be randomly added. If there are two or more objects to be recognized in a group, the average value for each attribute may be calculated.

[0027] Next, a probability distribution and an acceptable threshold for parameters for expanding or randomizing the target object are determined from the distribution of similarities and dissimilarities of each attribute within the group and outside the group (step S205).Furthermore, a probability distribution and an acceptable threshold for parameters for expanding or randomizing the background object are determined from the distribution of similarities and dissimilarities of each attribute within the group and outside the group (step S206).

[0028] Finally, learning data is generated based on the probability distribution and tolerance threshold of the target object determined in step S205 and the probability distribution and tolerance threshold of the background object determined in step S206 (step S207).

[0029] Example 1 Next, a method for generating the above-mentioned training data will be described based on a specific task with reference to FIGS. Task 1 shown in FIG. 4 is a task to classify multiple objects (five beverages in this example) by product. Specifically, product P1 and product P2 are classified into the same group GR1, product P3 into group GR2, product P4 into group GR3, and product P5 into group GR4. Note that although product names P1 and P2 are slightly different, "ABC Tea" and "New ABC Tea," their other attributes (size, color, shape, etc.) are almost the same, and so they are classified into the same group. In addition, the beverages in this example include products P1 to P3, which are 500 ml containers (relatively long containers in the vertical direction), and P4 and P5, which are 300 ml containers (relatively short containers in the vertical direction). Furthermore, products P1, P2, and P4 are light green. Product P3 is dark green. Product P5 is dark navy blue.

[0030] The following describes the scope of expansion or randomization when generating learning data for training a recognizer for group 3 (recognition target object P4). In this example, the resulting learning data is image data including the recognition target object (product P4 in group GR3) and background objects that are not to be recognized (see FIG. 5). In this example, the background objects are objects (products P1, P2, P3, and P5) in groups GR1, GR2, and GR4 other than group 3 (recognition target object P4).

[0031] Here, the similarity and the difference are calculated.

[0032] (1) Similarity of objects within group 3 The attribute "container length (vertical size)" is compared between multiple products in group 3. Here, group 3 only contains product P4, which is the object to be recognized, so the similarity can be determined to be high (i.e., 1). Also, the attribute "container color" is compared between multiple products in group 3. Here, group 3 only contains product P4, which is the object to be recognized, so the similarity can be determined to be high (i.e., 1).

[0033] (2) Differences between objects in Group 3 and other groups The attribute "container length (vertical size)" is compared between multiple products across different groups. Here, the degree of difference in container length between product P4, a recognition target object in group 3, and other objects in other groups (products P1, P2, P3, P5) is large (i.e., can be a value close to 1). Also, the attribute "container color" is compared between multiple products across different groups. Here, the degree of difference in color between product P4, a recognition target object in group 3, and other objects in other groups (products P1, P2, P3, P5) is large (i.e., can be a value close to 1).

[0034] The determination of the range of expansion or randomization based on the similarity and difference calculated for each attribute as described above will now be described. Here, the "range of expansion or randomization" refers to the degree to which the training data can be processed or the degree to which the simulation environment can be diversified. For example, the range of expansion or randomization can be expressed as a numerical value between 0 and 1. In this case, the more extensively an object can be processed, the closer the range of expansion or randomization will be to 1. Conversely, the more extensively an object cannot be processed, the closer the range of expansion or randomization will be to 0.

[0035] The expansion or randomization of the "target object" is determined based on the "similarity" described above. For the attribute "container length," the similarity is high (i.e., 1), as described above, so the randomization range is small (i.e., it can be any value close to 0 (e.g., 0.01)). Similarly, for the "container color," the similarity is high (i.e., 1), as described above, so the randomization range is small (i.e., it can be any value close to 0 (e.g., 0.01)). Thus, the first attribute (e.g., container length) of the target object is randomized within the set range (e.g., a range from 0 to 0.01), and the second attribute (e.g., container color) is expanded or randomized within the set range (e.g., a range from 0 to 0.01). In this way, for example, a "product with a relatively short vertical size and a light green container" (recognition target object P3-1 in FIG. 5) can be generated as training data.

[0036] Next, the randomization of the "background object" is determined based on the "dissimilarity" described above. For the attribute "container length," the dissimilarity is large (i.e., it can be a value close to 1 (e.g., 0.97)) as described above, so the range of expansion or randomization is large (i.e., it can be any value close to 1 (e.g., 0.97)). Therefore, background objects with various vertical sizes (lengths) can be generated within the range thus set (e.g., a range of 0 to 0.97). Similarly, for "color," the dissimilarity is large (i.e., it can be a value close to 1 (e.g., 0.90)) as described above, so the range of expansion or randomization is large (i.e., it can be any value close to 1 (e.g., 0.90)). Therefore, background objects with various colors (e.g., red, blue, green, etc.) can be generated within the range thus set (e.g., a range of 0 to 0.90). Therefore, the background objects can be expanded or randomized within the range of each attribute, and learning data can be generated, for example, such as "products in containers that are relatively short in vertical size and have a color other than green" (P1-2 in Figure 5) or "products in containers that are relatively long in vertical size and have a light green color" (P1-1 in Figure 5).

[0037] An example of training data generated by randomizing the target object and background objects as described above is shown in Figure 5. The generated training data IMG1 includes the recognition target product P3-1 (a modified version of P3) and two background objects P1-1 and P1-2 (a modified version of P1 with different colors and lengths).

[0038] <Example 2> Next, a method for generating the above-mentioned training data will be described based on a specific task with reference to FIGS. Task 2 shown in FIG. 6 is a task to classify multiple objects (five beverages in this example) by size. The beverages in this example include products P1 to P3, which are 500 ml containers (containers that are relatively long vertically), and products P4 and P5, which are 300 ml containers (containers that are relatively short vertically). Therefore, products P1, P2, and P3 are classified into the same group GR1, and products P4 and P5 are classified into group GR2. Note that although products P1 and P2 have slightly different product names, "ABC Tea" and "New ABC Tea," their other attributes (size, color, shape, etc.) are almost the same, and so they are classified into the same group. Furthermore, products P1, P2, and P4 are light green. Product P3 is dark green. Product P5 is dark navy blue.

[0039] The following describes the scope of expansion or randomization when generating training data for training the recognizers of group 1 (recognition target objects P1, P2, and P3) and group 2 (recognition target objects P4 and P5). In this example, the resulting training data is image data including the recognition target objects (products P4 and P5 in group GR2) and background objects that are not to be recognized (see FIG. 7). In this example, the background objects are the objects (products P1, P2, and P3) in group GR1 other than group 2 (recognition target objects P4 and P5).

[0040] Here, the similarity and difference are calculated as described above.

[0041] (1) Similarity of objects within group 1 The attribute "container length (vertical size)" is compared between multiple products in group 1. In this case, the recognition target objects products P1, P2, and P3 in group 1 are containers of approximately the same length, and the similarity can be very high (i.e., a value close to 1 (e.g., 0.99)). Also, the attribute "container color" is compared between multiple products in group 1. In this case, the recognition target objects products P1, P2, and P3 in group 1 have containers of various colors (light green and dark green), and the distribution of similarity can be quite high (i.e., a value close to 0 (e.g., 0.80)).

[0042] (2) Similarity of objects within group 2 The attribute "container length (vertical size)" is compared between multiple products in group 2. In this case, products P4 and P5 in group 2, which are objects to be recognized, are containers of approximately the same length, and the distribution of similarity can be determined to be high (i.e., 0.97). Furthermore, the attribute "container color" is compared between multiple products in group 2. In this case, the containers of products P4 and P5, which are objects to be recognized in group 2, are various colors (light green and dark navy blue), and the distribution of similarity can be determined to be quite small (i.e., 0.45).

[0043] (3) Difference between Group 2 and the other Group 1 objects The attribute "container length (vertical size)" is compared between multiple products across different groups. Here, the distribution of the degree of difference between the average value of the container length of products P4 and P5, which are objects to be recognized in group 2, and the average value of the container length of other objects in other groups (products P1, P2, P3), is relatively large (e.g., 0.75). Also, the attribute "container color" is compared between multiple products across different groups. Here, the distribution of the degree of difference between the average value of the container color of products P4 and P5, which are objects to be recognized in group 2 (e.g., a mixture of navy blue and light green), and the average value of the container color of other objects in other group 1 (products P1, P2, P3), which is a mixture of two light colors and light green, is somewhat large (e.g., 0.65).

[0044] The determination of the range of expansion or randomization based on the distribution of similarity and dissimilarity calculated for each attribute as described above will now be described.

[0045] (1) Expanding or randomizing the scope of objects in Group 1 The randomization of the "objects" in group 1 is determined based on the distributions of the "similarity" and "dissimilarity" described above. For the attribute "container length," the similarity is high (i.e., 0.99) as described above, so the randomization range is small (i.e., it can be any value close to 0 (e.g., 0.01)). Similarly, for the attribute "color," the similarity is high (i.e., 0.80) as described above, so the randomization range is small (i.e., it can be any value close to 0 (e.g., 0.20)). Thus, the objects in group 1 are randomized within the set range (e.g., 0 to 0.01) for the first attribute (e.g., container length), and within the set range (e.g., 0 to 0.20) for the second attribute (e.g., container color). Thus, for example, "products with relatively short vertical dimensions and green containers" can be generated as training data.

[0046] (2) Expanding or randomizing the scope of objects in Group 2 Next, the randomization of the "objects" in group 2 is determined based on the "similarity" and "dissimilarity" described above. For the attribute "container length," the similarity is large (i.e., a value close to 1 (e.g., 0.97)), as described above, so the range of expansion or randomization is small (i.e., any value close to 0 (e.g., 0.03)). Therefore, objects with varying vertical sizes (lengths) can be generated within this set range (e.g., a range from 0 to 0.03). For "color," the similarity is small (i.e., 0.45), as described above, so the range of expansion or randomization can be large (i.e., 0.45). Therefore, objects with various colors can be generated within this set range (e.g., a range from 0 to 0.45). Therefore, by randomizing the objects within this range for each attribute, it is possible to generate, for example, "products with containers that are relatively long vertically and have various colors other than green" as training data.

[0047] An example of training data generated by expanding or randomizing the target object and background objects as described above is shown in Figure 7. The generated training data includes product P3-1 (a processed version of P3) as the recognition target, and two background objects, products P1-1 and P1-2 (a different version of P1).

[0048] According to the second embodiment described above, the distribution of similarity and dissimilarity for each attribute can be obtained, and then the probability distribution and tolerance threshold of the expansion or randomization parameters can be determined.

[0049] 8 is a block diagram showing an example of the configuration of the information processing device 10 described in the above embodiment. Referring to FIG. 8, the information processing device 10 includes a processor 1201 and a memory 1202.

[0050] The processor 1201 reads and executes software (computer programs) from the memory 1202 to perform the processing of the information processing device 10 described using the flowcharts in the above-described embodiments. The processor 1201 may be, for example, a microprocessor, an MPU (Micro Processing Unit), or a CPU (Central Processing Unit). The processor 1201 may include multiple processors.

[0051] The memory 1202 is configured by a combination of volatile memory and non-volatile memory. The memory 1202 may include storage located remotely from the processor 1201. In this case, the processor 1201 may access the memory 1202 via an I / O interface (not shown).

[0052] 7, the memory 1202 is used to store a group of software modules. The processor 1201 reads and executes these software modules from the memory 1202, thereby performing the processing of the information processing device 10 described in the above embodiment.

[0053] As described with reference to FIG. 2, each of the processors included in the information processing device 10 executes one or more programs including a group of instructions for causing a computer to execute the algorithm described with reference to the drawing.

[0054] Although the above-described embodiments have been described as hardware configurations, the present disclosure is not limited to such configurations. Any processing in the present disclosure can also be realized by causing a processor to execute a computer program.

[0055] In the above examples, the program includes instructions (or software code) that, when loaded into a computer, cause the computer to perform one or more functions described in the embodiments. The program may be stored on a non-transitory computer-readable medium or a tangible storage medium. By way of example and not limitation, computer-readable medium or tangible storage medium includes random-access memory (RAM), read-only memory (ROM), flash memory, solid-state drive (SSD) or other memory technology, CD-ROM, digital versatile disc (DVD), Blu-ray® disc or other optical disk storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage device. The program may also be transmitted on a transitory computer-readable medium or communication medium. By way of example and not limitation, transitory computer-readable medium or communication medium includes electrical, optical, acoustic, or other forms of propagated signals.

[0056] The present invention is not limited to the above-described embodiment, and various modifications can be made without departing from the spirit and scope of the present invention. The above-described examples can also be implemented in combination with each other. [Explanation of symbols]

[0057] 10. Information processing equipment 11 Placement section 12 Grouping section 13 Area determination part 14 Range determination section 15 Generation part IMG1, IMG2 training data

Claims

1. an arrangement unit that arranges a plurality of objects in an n-dimensional state space based on n attributes; a grouping unit that groups the plurality of objects according to a target task; a similarity calculation unit that calculates a distribution of similarities of each attribute between each target object in the group; a dissimilarity calculation unit that calculates a distribution of dissimilarity of each attribute between one or more target objects to be recognized within the group and one or more objects outside the group; a target object range determination unit that determines a probability distribution and an allowable threshold of a randomization parameter of the target object from the distribution of the calculated similarity and dissimilarity for each attribute within the group and outside the group; a background object range determination unit that determines a probability distribution and an allowable threshold of randomization parameters of background objects that do not need to be recognized from the distribution of the calculated similarity and dissimilarity for each attribute within the group and outside the group; an information processing device comprising: a generation unit that generates learning data for object recognition by expanding or randomizing an object based on the determined probability distribution and tolerance threshold of the randomization parameters of the target object and the determined probability distribution and tolerance threshold of the randomization parameters of the background object.

2. The information processing apparatus according to claim 1 , further comprising a learning unit that performs machine learning using the generated object recognition learning data.

3. The information processing device according to claim 1 , wherein the attributes include at least one of color, length, shape, texture, material, and pattern.

4. 4. The information processing device according to claim 1, wherein the groups include a group consisting of one or more target objects and a group consisting of one or more background objects.

5. A plurality of objects are arranged in an n-dimensional state space with n attributes, grouping the plurality of objects according to a target task; Calculating a distribution of similarities of each attribute between each target object in the group; calculating a distribution of the degree of difference of each attribute between one or more target objects to be recognized within the group and one or more objects outside the group; determining a probability distribution and an acceptable threshold for a randomization parameter of the target object from the calculated distribution of the similarity and the dissimilarity for each attribute within the group and outside the group; determining a probability distribution and an acceptable threshold value of randomization parameters for background objects that do not need to be recognized from the distribution of the calculated similarity and dissimilarity for each attribute within the group and outside the group; Based on the determined probability distribution and tolerance threshold of the randomization parameters of the target object and the determined probability distribution and tolerance threshold of the randomization parameters of the background object, the object is expanded or randomized to generate learning data for object recognition. Information processing methods.

6. A plurality of objects are arranged in an n-dimensional state space with n attributes, grouping the plurality of objects according to a target task; Calculating a distribution of similarities of each attribute between each target object in the group; calculating a distribution of the degree of difference of each attribute between one or more target objects to be recognized within the group and one or more objects outside the group; determining a probability distribution and an acceptable threshold for a randomization parameter of the target object from the calculated distribution of the similarity and the dissimilarity for each attribute within the group and outside the group; determining a probability distribution and an acceptable threshold value of randomization parameters for background objects that do not need to be recognized from the distribution of the calculated similarity and dissimilarity for each attribute within the group and outside the group; A program that causes a computer to generate learning data for object recognition by expanding or randomizing an object based on the determined probability distribution and tolerance threshold of the randomization parameters of the target object and the determined probability distribution and tolerance threshold of the randomization parameters of the background object.

Images