Training device, training method, object recognition device, object recognition method, program and machine learning model

Abstract

To provide a prior learning technique of a machine learning model utilizing a non-real image.SOLUTION: One aspect of the present disclosure relates to a training device comprising: an acquisition unit which acquires FDSL (Formula-driven Supervised Learning) data representing a target object and CG (Computer Graphics) data; and a training unit which trains a machine learning model for training by utilizing the FDSL data and the CG data.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to a training device, a training method, an object recognition device, an object recognition method, a program, and a machine learning model. [Background technology]

[0002] When attempting to solve an individual task using a machine learning model, typically, training data for the task is prepared, and the machine learning model is trained using the prepared training data. Then, the trained machine learning model is used to perform inference processing on data to be inferred. Generally, training a machine learning model requires a large amount of training data, and the cost of obtaining the training data can be a problem.

[0003] To generate machine learning models for individual tasks more efficiently, approaches using pre-trained models have been proposed. It is known that, for example, in object recognition models implemented using deep learning models that recognize objects from images showing the objects, feature extraction is first performed on the input image, and then the object is identified based on the extracted features. This feature extraction part is common to all object recognition tasks regardless of the individual task, and only the identification part can be adjusted for each individual task. Therefore, a pre-trained model that can be used for all object recognition tasks can be generated, and the identification part of the pre-trained model can be adjusted for each individual task, i.e., by fine-tuning or transfer learning, an object recognition model suited to each individual task can be obtained. In this way, it is possible to generate an object recognition model for each individual task using a pre-trained model as a base model with a smaller amount of training data.

[0004] On the other hand, training such pre-trained models also requires a large amount of training data. In particular, the Vision Transformer (ViT), which has attracted attention in recent years, is known to have a greater impact on recognition accuracy than convolutional neural networks (CNNs) due to the amount of training data. For this reason, pre-training is performed using large-scale real image datasets collected from the web, such as ImageNet and JFT-300. However, these real image datasets have problems such as privacy issues, including facial images, bias in the number of labels, and inaccuracy of labels.

[0005] Using artificially synthesized datasets such as VisualAtom and VisDA can avoid privacy and label bias. The former is a dataset synthesized based on mathematical formulas using a technique called Formula-Driven Supervised Learning (FDSL). The latter is a dataset synthesized using CG technology and rendering based on 3D models of objects. While FDSL has achieved high recognition accuracy in various individual tasks, there are cases where its accuracy is inferior to that of pre-training with real images, and it has not yet been able to completely replace real images. [Prior art documents] [Non-patent literature]

[0006] [Non-Patent Document 1] Visual Atoms: Pre-training Vision Transformers with Sinusoidal Waves, Sora Takashima, Ryo Hayamizu, Nakamasa Inoue, Hirokatsu Kataoka and Rio Yokota, https: / / arxiv.org / abs / 2303.01112 Summary of the Invention [Problem to be solved by the invention]

[0007] The impact of differences in the datasets used for pre-training is also reflected in differences in the attention that serves as the basis for the trained models' decisions. When we checked the areas of focus of the models visualized using Attention Rollout, we found that the model trained on ImageNet-1k focused on a wide range of iconic parts of the object to be recognized, while the model trained on VisualAtom-1k focused on only a part of the object, failing to fully capture the features of the target object.

[0008] The inventors believe that the difference in attention between FDSL and ImageNet is due to the presence or absence of prior knowledge about general objects. FDSL is able to focus on the shape of an object by learning various contour information, but objects in the real world have more complex combinations of contours than the classes included in FDSL, and it is difficult to determine which object in an image to focus on without knowledge of general objects.

[0009] In view of the above problems, one object of the present disclosure is to provide a pre-training technique for a machine learning model using non-real images. [Means for solving the problem]

[0010] One aspect of the present disclosure relates to a training device having an acquisition unit that acquires Formula-driven Supervised Learning (FDSL) data and Computer Graphics (CG) data representing a target object, and a training unit that trains a machine learning model to be trained using the FDSL data and the CG data. [Effects of the Invention]

[0011] According to the present disclosure, it is possible to provide a pre-learning technique for a machine learning model that uses non-real images. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 is a schematic diagram illustrating a training process for a pre-trained model according to one embodiment of the present disclosure. [Figure 2] FIG. 2 is a schematic diagram illustrating a training device for training a pre-trained model according to one embodiment of the present disclosure, and an object recognition device using a trained object recognition model fine-tuned from the trained pre-trained model. [Figure 3] FIG. 3 is a block diagram illustrating the hardware configuration of a training device and an object recognition device according to an embodiment of the present disclosure. [Figure 4] FIG. 4 is a block diagram illustrating a functional configuration of a training device according to an embodiment of the present disclosure. [Figure 5] FIG. 5 is a block diagram illustrating a training process for a pre-trained model according to one embodiment of the present disclosure. [Figure 6] FIG. 6 is a diagram illustrating a calculation procedure for pre-training and fine-tuning according to an embodiment of the present disclosure. [Figure 7] FIG. 7 is a schematic diagram illustrating feature extractors and classifiers of a pre-trained model and a fine-tuned machine learning model according to one embodiment of the present disclosure. [Figure 8] FIG. 8 is a block diagram illustrating a training process for a pre-trained model according to one embodiment of the present disclosure. [Figure 9] FIG. 9 is a diagram illustrating a calculation procedure of pre-learning and additional pre-learning according to an embodiment of the present disclosure. [Figure 10] FIG. 10 is a schematic diagram illustrating feature extractors and classifiers of a pre-trained model and a fine-tuned machine learning model according to one embodiment of the present disclosure. [Figure 11] FIG. 11 is a flowchart illustrating a training process according to one embodiment of the present disclosure. [Figure 12] FIG. 12 is a block diagram illustrating a functional configuration of an object recognition device according to an embodiment of the present disclosure. [Figure 13] FIG. 13 is a flowchart illustrating an object recognition process according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings.

[0014] In the following examples, a training process is disclosed that produces a pretrained model.

[0015] [Summary of the Disclosure] To summarize the present disclosure, a training device 100 according to the following embodiment uses Formula-driven Supervised Learning (FDSL) data and Computer Graphics (CG) data representing a target object to train a machine learning model to be trained as a pre-learning model.

[0016] FIG. 1 is a schematic diagram illustrating a training process for a pre-trained model according to an embodiment of the present disclosure. For example, as shown in FIG. 1, a pre-trained model 30 to be trained is trained by supervised learning using a combination of FDSL data synthesized based on a mathematical formula and CG data synthesized by rendering based on a 3D model of an object. Here, the FDSL data may be generated according to any known FDSL technique, and the CG data may be generated using any known CG technique. For example, the FDSL data may be composed of an image of an object generated based on a mathematical formula and an FDSL class indicating the type of the object, and the CG data may be composed of a CG image of the object rendered and a CG class indicating the type of the object. Here, the FDSL class may or may not overlap with the CG class.

[0017] 2 is a schematic diagram illustrating a training device for training a pre-learned model according to an embodiment of the present disclosure, and an object recognition device using a trained object recognition model fine-tuned from the trained pre-learned model. As shown in FIG. 2, the training device 100 acquires FDSL data and CG data from a training data database (DB) 10 and trains a pre-learned model 30 to be trained.

[0018] Once the training of the pre-learning model 30 to be trained is completed, the trained pre-learning model 31 is fine-tuned for a specific task. For example, once an object recognition model 50 trained to recognize a specific object is acquired, the object recognition device 200 uses the trained object recognition model 50 to acquire an image of the recognition target, and outputs a recognition result for the object shown in the image.

[0019] The trained pre-learning model 31 obtained in this way learns contour information from FDSL data and object-likeness from CG data, learning both the contour information and object-likeness features. As a result, the trained pre-learning model 31 is expected to have a region of interest that extends to the entire object, compared to a pre-learning model trained only by FDSL. Therefore, a highly accurate pre-learning model can be generated by combining FDSL data artificially generated without using real images with CG data.

[0020] Here, the training device 100 and the object recognition device 200 are realized by a computing device such as a server, a personal computer, a smartphone, or a tablet, and may have a hardware configuration such as that shown in Fig. 3. That is, the training device 100 and the object recognition device 200 have a drive device 101, a storage device 102, a memory device 103, a processor 104, a user interface (UI) device 105, and a communication device 106, which are interconnected via a bus B.

[0021] The programs or instructions for realizing the various functions and processes described below in the training device 100 and the object recognition device 200 may be stored in a removable storage medium such as a CD-ROM (Compact Disk-Read Only Memory) or flash memory.

[0022] When the storage medium is set in the drive device 101, the program or instructions are installed from the storage medium to the storage device 102 or memory device 103 via the drive device 101. However, the program or instructions do not necessarily have to be installed from the storage medium, but may be downloaded from any external device via a network or the like.

[0023] The storage device 102 is realized by a hard disk drive or the like, and stores installed programs or instructions as well as files, data, etc. used to execute the programs or instructions.

[0024] The memory device 103 is realized by a random access memory, a static memory, or the like, and when a program or instruction is activated, reads and stores the program, instruction, data, or the like from the storage device 102. The storage device 102, the memory device 103, and the removable storage medium may be collectively referred to as a non-transitory storage medium.

[0025] The processor 104 may be realized by one or more CPUs (Central Processing Units), GPUs (Graphics Processing Units), processing circuitry, etc., which may be composed of one or more processor cores, and performs various functions and processes of the training device 100 and object recognition device 200 described below in accordance with programs, instructions, data such as parameters required to execute the programs or instructions, etc. stored in the memory device 103.

[0026] The user interface (UI) device 105 may be composed of input devices such as a keyboard, mouse, camera, microphone, etc., output devices such as a display, speaker, headset, printer, etc., and input / output devices such as a touch panel, and realizes an interface between the user and the training device 100 and the object recognition device 200. For example, the user operates the training device 100 and the object recognition device 200 by operating a GUI (Graphical User Interface) displayed on a display or touch panel using a keyboard, mouse, etc.

[0027] The communication device 106 is realized by various communication circuits that execute communication processes with external devices, the Internet, a communication network such as a LAN (Local Area Network), and the like.

[0028] However, the above-described hardware configuration is merely an example, and the training device 100 and the object recognition device 200 according to the present disclosure may be realized by any other appropriate hardware configuration.

[0029] [Training device] Next, a training device 100 according to an embodiment of the present disclosure will be described with reference to Fig. 4. The training device 100 according to this embodiment trains a pre-learned model 30 to be trained using artificially generated FDSL data and CG data.

[0030] 4 is a block diagram showing a functional configuration of a training device 100 according to one embodiment of the present disclosure. As shown in FIG. 4, the training device 100 includes an acquisition unit 110 and a training unit 120. For example, one or more functional units of the acquisition unit 110 and the training unit 120 may be realized by one or more processors 104 executing one or more programs or instructions stored in the memory device 103.

[0031] The acquisition unit 110 acquires FDSL data and CG data representing a target object. Specifically, the acquisition unit 110 accesses a training data DB 10 that stores FDSL data and CG data for training the pre-learning model 30 to be trained, and acquires the FDSL data and CG data for training from the training data DB 10.

[0032] For example, FDSL data is a synthetic image generated by mathematical processing such as affine transformation, and can be used to learn general features without using actual images. When FDSL is used to train ViT, methods specializing in contours (Radial Contour Database: RCDB) and methods diversifying contours using sinusoidal waves (VisualAtom) have been proposed, and the FDSL data in this embodiment can also be generated by any of these methods. The acquisition unit 110 acquires FDSL data consisting of an FDSL image and the class of the FDSL image, and passes the acquired FDSL data to the training unit 120 as training data for the pre-learning model 30 to be trained.

[0033] The CG data is a synthetic image generated by rendering based on a three-dimensional model of an object, and can be generated by randomly changing synthesis parameters such as lighting, texture, etc. The acquisition unit 110 acquires CG data consisting of a CG image and a class of the CG image, and passes the acquired CG data to the training unit 120 as training data for the pre-learning model 30 to be trained.

[0034] The training unit 120 trains the machine learning model to be trained using the FDSL data and the CG data. Specifically, the training unit 120 trains the pre-trained model 30 to be trained using a combination of the FDSL data and the CG data acquired from the acquisition unit 110. For example, the training unit 120 may train the pre-trained model 30 to be trained using a combination of the FDSL data and the CG data according to mixed learning or sequential learning, as described in detail below.

[0035] First, in mixed learning, the training unit 120 may generate a mixed image of FDSL data and CG data and train a machine learning model using the mixed image. That is, the training unit 120 may generate a mixed image of FDSL data and CG data by mixing the FDSL data and CG data, and train the pre-learning model 30 to be trained using the generated mixed image.

[0036] 5, when FDSL data and CG data are acquired, the training unit 120 generates a mixed image of the FDSL data and the CG data in step S11. For example, the training unit 120 generates a mixed image x as a linear sum of the FDSL data and the CG data according to the following equation: mix may be generated.

number

[0037] In this way, the mixed image x mix , the training unit 120 generates the mixed image x mix The pre-learning model 30 to be trained is trained by: When the pre-learning model 30 to be trained is composed of a feature extractor that extracts features of an object captured in an image and a classifier that identifies the object based on the extracted features, specifically, the training unit 120 may optimize the parameters of the pre-learning model 30 to be trained according to the following equation:

number

[0038] That is, the training unit 120 solves the classification problem for each of the FDSL data and the CG data, and uses the sum of these losses as the final loss to adjust parameters to minimize the sum of the losses. The trained pre-learning model 31 thus obtained learns contour information from the FDSL data and object-likeness from the CG data, thereby learning both the contour information and the object-likeness features. As a result, the trained pre-learning model 31 is expected to have a region of interest that extends to the entire object, compared to a pre-learning model trained only by FDSL.

[0039] Then, in step S13, the trained pre-learning model 31 is fine-tuned using actual images as downstream task data stored in the downstream task data DB13 to become a machine learning model suitable for an individual task, and a trained machine learning model 32 suitable for the task can be obtained.

[0040] The pre-learning in step S12 and the fine-tuning in step S13 described above can be specified, for example, by the calculation procedure shown in Fig. 6. In the calculation procedure for pre-learning shown in the figure, the mixture ratio α is set to 0.5, and when the pre-learning model 30 to be trained receives an input image of a mixture of FDSL data and CG data, it calculates the predicted probability φ of each class. θ,f (x mix ) and the training unit 120 outputs the parameter θ of the feature extractor and the parameter f of the classifier. fdsl and f cg and optimize.

[0041] On the other hand, in the illustrated fine-tuning computational procedure, the trained pre-trained model 31 is ft As an input, the predicted probability of each class φ θ,f (x ft ) Then, the training unit 120 optimizes the parameter θ of the feature extractor and the parameter f of the classifier.

[0042] In such fine-tuning, as shown in Figure 7, the feature extractor of the trained pre-learning model 31 is directly inherited by the fine-tuned machine learning model 32, and the discriminator f fdsl and f cg In a fine-tuned machine learning model, the discriminator f ft In other words, only the classifier is replaced.

[0043] Next, in sequential learning, the training unit 120 may generate an intermediate model by training a machine learning model using FDSL data, and train the intermediate model using CG data. That is, the training unit 120 first trains the pre-learning model 30 to be trained using FDSL data, and obtains the intermediate model 33 from the pre-learning model 30 to be trained. Next, the training unit 120 trains the intermediate model 33 using CG data, and obtains the trained pre-learning model 31 from the intermediate model 33.

[0044] 8, upon acquiring FDSL data and CG data, the training unit 120 trains (pre-learns) the pre-learning model 30 to be trained using the FDSL data in step S21. When the pre-learning model 30 to be trained includes a feature extractor that extracts features of an object captured in an image and a classifier that identifies the object based on the extracted features, specifically, the training unit 120 may optimize the parameters of the pre-learning model 30 to be trained according to the following equation:

number

[0045] After obtaining the intermediate model 33 in this manner, the training unit 120 then trains (additional pre-learning) the intermediate model 33 using CG data in step S22. Specifically, the training unit 120 may optimize the parameters of the intermediate model 33 according to the following equation:

number

[0046] The trained pre-learning model 31 obtained in this way learns contour information from FDSL data and object-likeness from CG data, learning both contour information and object-likeness characteristics. As a result, the trained pre-learning model 31 is expected to have a wider area of ​​interest for the entire object compared to a pre-learning model trained only with FDSL. Furthermore, the model's tendency to focus only on partial contours in FDSL is weakened, and it is expected to be able to capture the entire object better than mixed learning.

[0047] Then, in step S23, the trained pre-learning model 31 is fine-tuned using actual images as downstream task data stored in the downstream task data DB13 to become a machine learning model 32 suitable for an individual task, and a trained machine learning model 32 suitable for the task can be obtained.

[0048] The above-mentioned pre-learning in step S21 and the additional pre-learning in step S22 can be defined as, for example, a calculation procedure as shown in Fig. 9. In the calculation procedure of the pre-learning shown in the figure, when the pre-learning model 30 to be trained receives FDSL data as input, it calculates the predicted probability φ of each class. θ,f (x fdsl ) Then, the training unit 120 optimizes the parameter θ of the feature extractor and the parameter f of the classifier.

[0049] In the calculation procedure of the additional pre-learning shown in the figure, when the intermediate model 33 receives CG data as input, it calculates the predicted probability φ θ,f (x cg ) Then, the training unit 120 optimizes the parameter θ of the feature extractor and the parameter f of the classifier.

[0050] The fine tuning in step S23 shares the same calculation procedure as the fine tuning in step S13.

[0051] In such fine-tuning, the feature extractor of the trained pre-trained model 31 is directly passed on to the fine-tuned machine learning model 32, and the discriminator f cg In a fine-tuned machine learning model, the discriminator f ft In other words, only the classifier is replaced. The feature extractor of the intermediate model 33 trained with the FDSL data is directly inherited by the trained pre-learning model 31 trained with the CG data, and the classifier f fdsl In the trained pre-learning model 31, the discriminator f cgThis becomes:

[0052] [Training process] Next, a training process for the pre-learned model 30 to be trained according to an embodiment of the present disclosure will be described. Fig. 11 is a flowchart showing the training process for the pre-learned model 30 to be trained according to an embodiment of the present disclosure. The training process is performed by the training device 100 described above, and more specifically, may be realized by one or more processors 104 of the training device 100 executing one or more programs or instructions stored in one or more memory devices 103.

[0053] 11, in step S101, the training device 100 acquires FDSL data and CG data. Specifically, the training device 100 acquires the FDSL data and CG data from the training data DB 10 as training data for pre-learning.

[0054] In step S102, the training device 100 trains a machine learning model to be trained using the FDSL data and the CG data. Specifically, the training device 100 uses both the FDSL and the CG data to train a pre-learning model to be trained by mixed learning or sequential learning.

[0055] In mixed learning, the training device 100 generates a mixed image of FDSL data and CG data, and trains the pre-trained model 30 to be trained using the generated mixed image. On the other hand, in sequential learning, the training device 100 first trains the pre-trained model 30 to be trained using FDSL data, and then trains the intermediate model 33 obtained by the training using CG data.

[0056] The trained pre-learning model 31 obtained in this manner can be fine-tuned using training data such as real images corresponding to individual tasks, thereby obtaining a trained machine learning model 32 suitable for the task.

[0057] According to the training device 100 and training process described above, a highly accurate pre-learning model can be generated by using FDSL data and CG data that are artificially generated without using real images.

[0058] [Object recognition device] Next, an object recognition device 200 according to an embodiment of the present disclosure will be described. The object recognition device 200 according to this embodiment uses a trained object recognition model 50 that has been fine-tuned for an object recognition task from a pre-learning model 31 trained by the training device 100, and outputs a recognition result for an object captured in an image to be recognized.

[0059] 12 is a block diagram showing a functional configuration of an object recognition device 200 according to an embodiment of the present disclosure. As shown in FIG. 12, the object recognition device 200 includes an acquisition unit 210 and a recognition unit 220. For example, one or more functional units of the acquisition unit 210 and the recognition unit 220 may be realized by one or more processors 104 executing one or more programs or instructions stored in the memory device 103.

[0060] The acquisition unit 210 acquires an image showing an object to be recognized. For example, the acquisition unit 210 acquires an image showing an object to be recognized captured by a camera, and passes the acquired image to the recognition unit 220.

[0061] The recognition unit 220 outputs an object recognition result using a trained machine learning model that has been trained using FDSL data and CG data representing the target object. Specifically, the recognition unit 220 inputs an image showing the object to be recognized from the pre-learning model 31 trained by the training device 100 to the trained object recognition model 50 that has been fine-tuned for the object recognition task, and obtains and outputs the recognition result of the object from the trained object recognition model 50.

[0062] Here, the object recognition model 50 is a machine learning model that causes a computer to acquire images of objects to be recognized and output object recognition results. The object recognition model 50 is trained using FDSL data and CG data representing the target objects. In other words, the object recognition model 50 can be generated by fine-tuning the trained pre-learning model 31 described above using training data consisting of actual images of the objects to be recognized and the classes of the objects.

[0063] [Object recognition processing] Next, an object recognition process using the trained object recognition model 50 according to an embodiment of the present disclosure will be described. Fig. 13 is a flowchart showing the object recognition process using the trained object recognition model 50 according to an embodiment of the present disclosure. The object recognition process is performed by the above-described object recognition device 200, and more specifically, may be realized by one or more processors 104 of the object recognition device 200 executing one or more programs or instructions stored in one or more memory devices 103.

[0064] 13, in step S201, the object recognition device 200 acquires an image showing an object to be recognized. For example, the object recognition device 200 may receive an image capturing a specific view including the object from a camera such as a surveillance camera installed at a predetermined position via a wireless connection / wired connection.

[0065] In step S202, the object recognition device 200 outputs an object recognition result using the object recognition model 50. For example, when an image of a specific view is received from a surveillance camera, the object recognition device 200 inputs the received image into the object recognition model 50 and obtains a recognition result indicating the class of the object shown in the image from the object recognition model 50. The object recognition device 200 may then perform a predetermined action based on the obtained recognition result. For example, when the object recognized in the image is detected to belong to a predetermined class such as a dangerous object, the object recognition device 200 may notify an operator or the like that a dangerous object has been observed.

[0066] According to the object recognition device 200 and object recognition processing described above, by utilizing an object recognition model 50 that is fine-tuned for an individual task from a pre-learning model 31 that is generated using a combination of FDSL data that is artificially generated without using real images and CG data, it is possible to achieve more accurate object recognition processing without preparing a large amount of training data for that task.

[0067] (Appendix 1) an acquisition unit that acquires Formula-driven Supervised Learning (FDSL) data and Computer Graphics (CG) data representing a target object; a training unit that trains a machine learning model to be trained using the FDSL data and the CG data; A training device having: (Appendix 2) 2. The training device of claim 1, wherein the training unit generates a mixed image of the FDSL data and the CG data and trains the machine learning model using the mixed image. (Appendix 3) 3. The training device of claim 2, wherein the mixed image is generated as a linear sum of the FDSL data and the CG data. (Appendix 4) The training department: generating an intermediate model by training the machine learning model using the FDSL data; The training device according to any one of appendices 1 to 3, wherein the intermediate model is trained using the CG data. (Appendix 5) 5. The training device according to any one of claims 1 to 4, wherein the machine learning model is a pre-trained model. (Appendix 6) Acquiring Formula-driven Supervised Learning (FDSL) data and Computer Graphics (CG) data representing a target object; training a machine learning model to be trained using the FDSL data and the CG data; A training method in which the computer executes the above. (Appendix 7) Acquiring Formula-driven Supervised Learning (FDSL) data and Computer Graphics (CG) data representing a target object; training a machine learning model to be trained using the FDSL data and the CG data; A program that causes a computer to execute the following. (Appendix 8) an acquisition unit that acquires an image showing an object to be recognized; a recognition unit that uses a machine learning model to output a recognition result of the object; and The object recognition device, wherein the machine learning model is trained using Formula-driven Supervised Learning (FDSL) data and Computer Graphics (CG) data representing a target object. (Appendix 9) acquiring an image showing an object to be recognized; outputting a recognition result of the object using a machine learning model; The computer executes An object recognition method in which the machine learning model is trained using Formula-driven Supervised Learning (FDSL) data and Computer Graphics (CG) data representing a target object. (Appendix 10) acquiring an image showing an object to be recognized; outputting a recognition result of the object using a machine learning model; The computer is running, A program in which the machine learning model is trained using Formula-driven Supervised Learning (FDSL) data and Computer Graphics (CG) data representing a target object. (Appendix 11) acquiring an image showing an object to be recognized; outputting the object recognition result; A machine learning model that causes a computer to execute The machine learning model is trained using Formula-driven Supervised Learning (FDSL) data and Computer Graphics (CG) data representing a target object.

[0068] Although the examples of the present disclosure have been described in detail above, the present disclosure is not limited to the specific embodiments described above, and various modifications and variations are possible within the scope of the gist of the present disclosure as set forth in the claims. [Industrial Applicability]

[0069] This disclosure is useful for pre-trained models. [Explanation of symbols]

[0070] 10 Training Data DB 30 Pre-trained model to be trained 50 pre-trained object recognition models 100 training equipment 110 Acquisition Department 120 Training Department 200 Object recognition device 210 Acquisition Department 220 Recognition part

Claims

1. an acquisition unit that acquires Formula-driven Supervised Learning (FDSL) data and Computer Graphics (CG) data representing a target object; a training unit that uses the FDSL data and the CG data to train a machine learning model to be trained; A training device having:

2. The training device according to claim 1 , wherein the training unit generates a mixed image of the FDSL data and the CG data, and trains the machine learning model using the mixed image.

3. The training device of claim 2 , wherein the blended image is generated as a linear sum of the FDSL data and the CG data.

4. The training department: generating an intermediate model by training the machine learning model using the FDSL data; The training device of claim 1 , wherein the CG data is utilized to train the intermediate model.

5. The training device of claim 1 , wherein the machine learning model is a pre-trained model.

6. Acquiring Formula-driven Supervised Learning (FDSL) data and Computer Graphics (CG) data representing a target object; training a machine learning model to be trained using the FDSL data and the CG data; A training method in which the computer executes the above.

7. Acquiring Formula-driven Supervised Learning (FDSL) data and Computer Graphics (CG) data representing a target object; training a machine learning model to be trained using the FDSL data and the CG data; A program that causes a computer to execute the following.

8. an acquisition unit that acquires an image showing an object to be recognized; a recognition unit that uses a machine learning model to output a recognition result of the object; and The object recognition device, wherein the machine learning model is trained using Formula-driven Supervised Learning (FDSL) data and Computer Graphics (CG) data representing a target object.

9. acquiring an image showing an object to be recognized; outputting a recognition result of the object using a machine learning model; The computer executes The object recognition method, wherein the machine learning model is trained using Formula-driven Supervised Learning (FDSL) data and Computer Graphics (CG) data representing the target object.

10. acquiring an image showing an object to be recognized; outputting a recognition result of the object using a machine learning model; The computer is running, The machine learning model is trained using Formula-driven Supervised Learning (FDSL) data and Computer Graphics (CG) data representing a target object.

11. acquiring an image showing an object to be recognized; outputting the object recognition result; A machine learning model that causes a computer to execute The machine learning model is trained using Formula-driven Supervised Learning (FDSL) data and Computer Graphics (CG) data representing a target object.

Images