Patent
Atlas literature
Patent
US 10,692,002Patent drawings and their descriptions. Click a drawing to enlarge it.
Claims define the patent's legal scope. Independent claims stand alone; dependent claims (nested) narrow them. Click a claim to expand its dependents.
A method for learning a pedestrian detector to be used for robust surveillance based on image analysis, comprising steps of: (a) a learning device, if at least one training image with each of bounding boxes corresponding to each of pedestrians for training generated thereon is acquired, generating one or more image patches by cropping each of regions, corresponding to each of the bounding boxes, on the training image, and instructing an adversarial style transformer to generate one or more transformed image patches by converting each of the pedestrians for training corresponding to each of the image patches into each of transformed pedestrians capable of impeding a detection by the pedestrian detector; and (b) the learning device generating at least one transformed training image by replacing each of the regions, corresponding to each of the bounding boxes, on the training image with the transformed image patches, and learning the pedestrian detector by usin g the transformed ima g e patches in which the pedestrians are more dif fi cult to detect by the pedestrian detector than the one or more image patches from the training image, wherein the step of learning the pedestrian detector includes (i) instructing the pedestrian detector to generate pedestrian detection information for training by detecting the transformed pedestrians located in the transformed training image, (ii) instructing a fi rst loss layer to calculate one or more fi rst losses by referring to each piece of the pedestrian detection information for training and its corresponding GT (Ground Truth), and (iii) learning at least part of one or more parameters of the pedestrian detector to minimize the fi rst losses, wherein, at the step of (a), the learning device resizes at least one of the image patches such that the image patches have a same size, and instructs the adversarial style transformer to output the transformed image patches with the same size, and wherein, at the step of (b), the learning device resizes the transformed image patches with the same size such that each of the transformed image patches has its own corresponding original size before the transformation, and then generates the transformed training image. Currently amended
The method of Claim 1, wherein, at the step of (b), the learning device instructs a discriminator to generate each of pedestrian scores for training representing each of probabilities of each of the transformed image patches being each of the pedestrians for training, and to further learn at least part of one or more parameters of the adversarial style transformer to maximize the pedestrian scores for training and the first losses. Original
The method of Claim 1, wherein the adversarial style transformer includes an encoder having one or more convolutional layers for applying one or more convolution operations to the image patches, and a decoder having one or more deconvolutional layers for applying one or more deconvolution operations to at least one feature map outputted from the encoder. Original
The method of Claim 1, wherein the training image is acquired from one of (i) the pedestrian detector outputting pedestrian detection information for testing including the bounding boxes by detecting pedestrians for testing located in at least one test image acquired from a surveillance camera, and (ii) a database storing the training image with its own true labels corresponding to the bounding boxes. Original
Canceled
A testing method for a pedestrian detector for testing to be used for robust surveillance based on image analysis, comprising steps of (a) a testing device, on condition that (1) a learning device has generated one or more image patches for training by cropping each of regions for training, corresponding to each of the bounding boxes for training, on at least one training image, which includes each of the bounding boxes for training corresponding to each of pedestrians for training generated thereon, and has instructed an adversarial style transformer to generate one or more transformed image patches for training by converting each of the pedestrians for training corresponding to each of the image patches for training into each of transformed pedestrians for training capable of impeding a detection for training by the pedestrian detector, and (2) the learning device has been generated at least one transformed training image by replacing each of the regions for training, corresponding to each of the bounding boxes for training, on the training image with the transformed image patches for training, has instructed the pedestrian detector to generate pedestrian detection information for training by detecting the transformed pedestrians for training located in the transformed training image, has instructed a first loss layer to calculate one or more first losses by referring to each piece of the pedestrian detection information for training and its corresponding GT (Ground Truth), and has learned at least part of one or more parameters of the pedestrian detector to minimize the first losses; if at least one test image with each of bounding boxes for testing corresponding to each of pedestrians for testing generated thereon is acquired, generating one or more image patches for testing by cropping each of regions for testing, corresponding to each of the bounding boxes for testing, on the test image, and instructing an adversarial style transformer to generate one or more transformed image patches for testing by converting each of the pedestrians for testing corresponding to each of the image patches for testing into each of transformed pedestrians for testing capable of impeding a detection for testing by the pedestrian detector; and (b) the testing device generating at least one transformed test image by replacing each of the regions for testing, corresponding to each of the bounding boxes for testing, on the test image with the transformed image patches for testing, and learning the pedestrian detector by using the transformed image patches in which the pedestrians are more difficult to detect by the pedestrian detector than the one or more ima g e patches from the trainin g image, wherein the lea ming the pedestrian detector includes [[and]] instructing the pedestrian detector to generate pedestrian detection information for testing by detecting the transformed pedestrians for testing located in the transformed test image; wherein, at the step of (a), the learning device resizes at least one of the image patches such that the image patches have a same size, and instructs the adversarial style transformer to output the transformed image patches with the same size, and wherein, at the step of (b), the learning device resizes the transformed image patches with the same size such that each of the transformed image patches has its own corresponding original size before the transformation, and then generates the transformed training image. Currently amended
The testing method of Claim 8, wherein, at the step of (b), the testing device instructs a discriminator to generate each of pedestrian scores for testing representing each of probabilities of each of the transformed image patches for testing being each of the pedestrians for testing. Original
The testing method of Claim 8, wherein the adversarial style transformer includes an encoder having one or more convolutional layers for applying one or more convolution operations to the image patches for testing, and a decoder having one or more deconvolutional layers for applying one or more deconvolution operations to at least one feature map for testing outputted from the encoder. Original
The testing method of Claim 8, wherein the training image has been acquired from one of (i) the pedestrian detector outputting the pedestrian detection information for testing including the bounding boxes for testing by detecting the pedestrians for testing located in the test image acquired from a surveillance camera, and (ii) a database storing the training image with its own true labels corresponding to the bounding boxes for training. Original
Canceled
A learning device for learning a pedestrian detector to be used for robust surveillance based on image analysis, comprising: at least one memory that stores instructions; and at least one processor configured to execute the instructions to: perform processes of (I) if at least one training image with each of bounding boxes corresponding to each of pedestrians for training generated thereon is acquired, generating one or more image patches by cropping each of regions, corresponding to each of the bounding boxes, on the training image, and instructing an adversarial style transformer to generate one or more transformed image patches by converting each of the pedestrians for training corresponding to each of the image patches into each of transformed pedestrians capable of impeding a detection by the pedestrian detector, and (II) generating at least one transformed training image by replacing each of the regions, corresponding to each of the bounding boxes, on the training image with the transformed image patches, and learning the pedestrian detector by using the transformed image patches in which the pedestrians are more dif fi cult to detect by the pedestrian detector than the one or more image patches from the training image, wherein the learning the pedestrian detector includes (i) instructing the pedestrian detector to generate pedestrian detection information for training by detecting the transformed pedestrians located in the transformed training image, (ii) instructing a fi rst loss layer to calculate one or more fi rst losses by referring to each piece of the pedestrian detection information for training and its corresponding GT (Ground Truth), and (iii) lea rn ing at least part of one or more parameters of the pedestrian detector to minimize the fi rst losses, wherein, at the process of (I), the learning device resizes at least one of the image patches such that the image patches have a same size, and instructs the adversarial style transformer to output the transformed image patches with the same size, and wherein, at the process of (II), the learning device resizes the transformed image patches with the same size such that each of the transformed image patches has its own corresponding original size before the transformation, and then generates the transformed training image. Currently amended
The learning device of Claim 15, wherein, at the process of (II), the processor instructs a discriminator to generate each of pedestrian scores for training representing each of probabilities of each of the transformed image patches being each of the pedestrians for training, and to further learn at least part of one or more parameters of the adversarial style transformer to maximize the pedestrian scores for training and the first losses. Original
The learning device of Claim 15, wherein the adversarial style transformer includes an encoder having one or more convolutional layers for applying one or more convolution operations to the image patches, and a decoder having one or more deconvolutional layers for applying one or more deconvolution operations to at least one feature map outputted from the encoder. Original
The learning device of Claim 15, wherein the training image is acquired from one of (i) the pedestrian detector outputting pedestrian detection information for testing including the bounding boxes by detecting pedestrians for testing located in at least one test image acquired from a surveillance camera, and (ii) a database storing the training image with its own true labels corresponding to the bounding boxes. Original
Canceled
A testing device for a pedestrian detector for testing to be used for robust surveillance based on image analysis, comprising: at least one memory that stores instructions; and at least one processor, on condition that (1) a learning device has generated one or more image patches for training by cropping each of regions for training, corresponding to each of the bounding boxes for training, on at least one training image, which includes each of the bounding boxes for training corresponding to each of pedestrians for training generated thereon, and has instructed an adversarial style transformer to generate one or more transformed image patches for training by converting each of the pedestrians for training corresponding to each of the image patches for training into each of transformed pedestrians for training capable of impeding a detection for training by the pedestrian detector, and (2) the learning device has [[been]] generated at least one transformed training image by replacing each of the regions for training, corresponding to each of the bounding boxes for training, on the training image with the transformed image patches for training, and has learned the pedestrian detector by usin g the transformed image patches in which the pedestrians are more difficult to detect by the pedestrian detector than the one or more image patches from the training image, wherein the learning of the pedestrian detector included (i) instructing the pedestrian detector to generate pedestrian detection information for training by detecting the transformed pedestrians for training located in the transformed training image, (ii) instructing a first loss layer to calculate one or more first losses by referring to each piece of the pedestrian detection information for training and its corresponding GT (Ground Truth), and (iii) lea rn ing at least part of one or more parameters of the pedestrian detector to minimize the first losses; configured to execute the instructions to: perform processes of: (I) if at least one test image with each of bounding boxes for testing corresponding to each of pedestrians for testing generated thereon is acquired, generating one or more image patches for testing by cropping each of regions for testing, corresponding to each of the bounding boxes for testing, on the test image, and instructing an adversarial style transformer to generate one or more transformed image patches for testing by converting each of the pedestrians for testing corresponding to each of the image patches for testing into each of transformed pedestrians for testing capable of impeding a detection for testing by the pedestrian detector, and (II) generating at least one transformed test image by replacing each of the regions for testing, corresponding to each of the bounding boxes for testing, on the test image with the transformed image patches for testing, and instructing the pedestrian detector to generate pedestrian detection information for testing by detecting the transformed pedestrians for testing located in the transformed test image, wherein, at the process of (I), the processor resizes at least one of the image patches for testing such that the image patches for testing have a same size, and instructs the adversarial style transformer to output the transformed image patches for testing with the same size, and wherein, at the process of (II), the processor resizes the transformed image patches for testing with the same size such that each of the transformed image patches for testing has its own corresponding original size before the transformation, and then generates the transformed test image. Currently amended
The testing device of Claim 22, wherein, at the process of (II), the processor instructs a discriminator to generate each of pedestrian scores for testing representing each of probabilities of each of the transformed image patches for testing being each of the pedestrians for testing. Original
The testing device of Claim 22, wherein the adversarial style transformer includes an encoder having one or more convolutional layers for applying one or more convolution operations to the image patches for testing, and a decoder having one or more deconvolutional layers for applying one or more deconvolution operations to at least one feature map for testing outputted from the encoder. Original
The testing device of Claim 22, wherein the training image has been acquired from one of (i) the pedestrian detector outputting the pedestrian detection information for testing including the bounding boxes for testing by detecting the pedestrians for testing located in the test image acquired from a surveillance camera, and (ii) a database storing the training image with its own true labels corresponding to the bounding boxes for training. Original
Canceled
Patent
Atlas literature
Patent
US 10,692,002Patent drawings and their descriptions. Click a drawing to enlarge it.
Claims define the patent's legal scope. Independent claims stand alone; dependent claims (nested) narrow them. Click a claim to expand its dependents.
A method for learning a pedestrian detector to be used for robust surveillance based on image analysis, comprising steps of: (a) a learning device, if at least one training image with each of bounding boxes corresponding to each of pedestrians for training generated thereon is acquired, generating one or more image patches by cropping each of regions, corresponding to each of the bounding boxes, on the training image, and instructing an adversarial style transformer to generate one or more transformed image patches by converting each of the pedestrians for training corresponding to each of the image patches into each of transformed pedestrians capable of impeding a detection by the pedestrian detector; and (b) the learning device generating at least one transformed training image by replacing each of the regions, corresponding to each of the bounding boxes, on the training image with the transformed image patches, and learning the pedestrian detector by usin g the transformed ima g e patches in which the pedestrians are more dif fi cult to detect by the pedestrian detector than the one or more image patches from the training image, wherein the step of learning the pedestrian detector includes (i) instructing the pedestrian detector to generate pedestrian detection information for training by detecting the transformed pedestrians located in the transformed training image, (ii) instructing a fi rst loss layer to calculate one or more fi rst losses by referring to each piece of the pedestrian detection information for training and its corresponding GT (Ground Truth), and (iii) learning at least part of one or more parameters of the pedestrian detector to minimize the fi rst losses, wherein, at the step of (a), the learning device resizes at least one of the image patches such that the image patches have a same size, and instructs the adversarial style transformer to output the transformed image patches with the same size, and wherein, at the step of (b), the learning device resizes the transformed image patches with the same size such that each of the transformed image patches has its own corresponding original size before the transformation, and then generates the transformed training image. Currently amended
The method of Claim 1, wherein, at the step of (b), the learning device instructs a discriminator to generate each of pedestrian scores for training representing each of probabilities of each of the transformed image patches being each of the pedestrians for training, and to further learn at least part of one or more parameters of the adversarial style transformer to maximize the pedestrian scores for training and the first losses. Original
The method of Claim 1, wherein the adversarial style transformer includes an encoder having one or more convolutional layers for applying one or more convolution operations to the image patches, and a decoder having one or more deconvolutional layers for applying one or more deconvolution operations to at least one feature map outputted from the encoder. Original
The method of Claim 1, wherein the training image is acquired from one of (i) the pedestrian detector outputting pedestrian detection information for testing including the bounding boxes by detecting pedestrians for testing located in at least one test image acquired from a surveillance camera, and (ii) a database storing the training image with its own true labels corresponding to the bounding boxes. Original
Canceled
A testing method for a pedestrian detector for testing to be used for robust surveillance based on image analysis, comprising steps of (a) a testing device, on condition that (1) a learning device has generated one or more image patches for training by cropping each of regions for training, corresponding to each of the bounding boxes for training, on at least one training image, which includes each of the bounding boxes for training corresponding to each of pedestrians for training generated thereon, and has instructed an adversarial style transformer to generate one or more transformed image patches for training by converting each of the pedestrians for training corresponding to each of the image patches for training into each of transformed pedestrians for training capable of impeding a detection for training by the pedestrian detector, and (2) the learning device has been generated at least one transformed training image by replacing each of the regions for training, corresponding to each of the bounding boxes for training, on the training image with the transformed image patches for training, has instructed the pedestrian detector to generate pedestrian detection information for training by detecting the transformed pedestrians for training located in the transformed training image, has instructed a first loss layer to calculate one or more first losses by referring to each piece of the pedestrian detection information for training and its corresponding GT (Ground Truth), and has learned at least part of one or more parameters of the pedestrian detector to minimize the first losses; if at least one test image with each of bounding boxes for testing corresponding to each of pedestrians for testing generated thereon is acquired, generating one or more image patches for testing by cropping each of regions for testing, corresponding to each of the bounding boxes for testing, on the test image, and instructing an adversarial style transformer to generate one or more transformed image patches for testing by converting each of the pedestrians for testing corresponding to each of the image patches for testing into each of transformed pedestrians for testing capable of impeding a detection for testing by the pedestrian detector; and (b) the testing device generating at least one transformed test image by replacing each of the regions for testing, corresponding to each of the bounding boxes for testing, on the test image with the transformed image patches for testing, and learning the pedestrian detector by using the transformed image patches in which the pedestrians are more difficult to detect by the pedestrian detector than the one or more ima g e patches from the trainin g image, wherein the lea ming the pedestrian detector includes [[and]] instructing the pedestrian detector to generate pedestrian detection information for testing by detecting the transformed pedestrians for testing located in the transformed test image; wherein, at the step of (a), the learning device resizes at least one of the image patches such that the image patches have a same size, and instructs the adversarial style transformer to output the transformed image patches with the same size, and wherein, at the step of (b), the learning device resizes the transformed image patches with the same size such that each of the transformed image patches has its own corresponding original size before the transformation, and then generates the transformed training image. Currently amended
The testing method of Claim 8, wherein, at the step of (b), the testing device instructs a discriminator to generate each of pedestrian scores for testing representing each of probabilities of each of the transformed image patches for testing being each of the pedestrians for testing. Original
The testing method of Claim 8, wherein the adversarial style transformer includes an encoder having one or more convolutional layers for applying one or more convolution operations to the image patches for testing, and a decoder having one or more deconvolutional layers for applying one or more deconvolution operations to at least one feature map for testing outputted from the encoder. Original
The testing method of Claim 8, wherein the training image has been acquired from one of (i) the pedestrian detector outputting the pedestrian detection information for testing including the bounding boxes for testing by detecting the pedestrians for testing located in the test image acquired from a surveillance camera, and (ii) a database storing the training image with its own true labels corresponding to the bounding boxes for training. Original
Canceled
A learning device for learning a pedestrian detector to be used for robust surveillance based on image analysis, comprising: at least one memory that stores instructions; and at least one processor configured to execute the instructions to: perform processes of (I) if at least one training image with each of bounding boxes corresponding to each of pedestrians for training generated thereon is acquired, generating one or more image patches by cropping each of regions, corresponding to each of the bounding boxes, on the training image, and instructing an adversarial style transformer to generate one or more transformed image patches by converting each of the pedestrians for training corresponding to each of the image patches into each of transformed pedestrians capable of impeding a detection by the pedestrian detector, and (II) generating at least one transformed training image by replacing each of the regions, corresponding to each of the bounding boxes, on the training image with the transformed image patches, and learning the pedestrian detector by using the transformed image patches in which the pedestrians are more dif fi cult to detect by the pedestrian detector than the one or more image patches from the training image, wherein the learning the pedestrian detector includes (i) instructing the pedestrian detector to generate pedestrian detection information for training by detecting the transformed pedestrians located in the transformed training image, (ii) instructing a fi rst loss layer to calculate one or more fi rst losses by referring to each piece of the pedestrian detection information for training and its corresponding GT (Ground Truth), and (iii) lea rn ing at least part of one or more parameters of the pedestrian detector to minimize the fi rst losses, wherein, at the process of (I), the learning device resizes at least one of the image patches such that the image patches have a same size, and instructs the adversarial style transformer to output the transformed image patches with the same size, and wherein, at the process of (II), the learning device resizes the transformed image patches with the same size such that each of the transformed image patches has its own corresponding original size before the transformation, and then generates the transformed training image. Currently amended
The learning device of Claim 15, wherein, at the process of (II), the processor instructs a discriminator to generate each of pedestrian scores for training representing each of probabilities of each of the transformed image patches being each of the pedestrians for training, and to further learn at least part of one or more parameters of the adversarial style transformer to maximize the pedestrian scores for training and the first losses. Original
The learning device of Claim 15, wherein the adversarial style transformer includes an encoder having one or more convolutional layers for applying one or more convolution operations to the image patches, and a decoder having one or more deconvolutional layers for applying one or more deconvolution operations to at least one feature map outputted from the encoder. Original
The learning device of Claim 15, wherein the training image is acquired from one of (i) the pedestrian detector outputting pedestrian detection information for testing including the bounding boxes by detecting pedestrians for testing located in at least one test image acquired from a surveillance camera, and (ii) a database storing the training image with its own true labels corresponding to the bounding boxes. Original
Canceled
A testing device for a pedestrian detector for testing to be used for robust surveillance based on image analysis, comprising: at least one memory that stores instructions; and at least one processor, on condition that (1) a learning device has generated one or more image patches for training by cropping each of regions for training, corresponding to each of the bounding boxes for training, on at least one training image, which includes each of the bounding boxes for training corresponding to each of pedestrians for training generated thereon, and has instructed an adversarial style transformer to generate one or more transformed image patches for training by converting each of the pedestrians for training corresponding to each of the image patches for training into each of transformed pedestrians for training capable of impeding a detection for training by the pedestrian detector, and (2) the learning device has [[been]] generated at least one transformed training image by replacing each of the regions for training, corresponding to each of the bounding boxes for training, on the training image with the transformed image patches for training, and has learned the pedestrian detector by usin g the transformed image patches in which the pedestrians are more difficult to detect by the pedestrian detector than the one or more image patches from the training image, wherein the learning of the pedestrian detector included (i) instructing the pedestrian detector to generate pedestrian detection information for training by detecting the transformed pedestrians for training located in the transformed training image, (ii) instructing a first loss layer to calculate one or more first losses by referring to each piece of the pedestrian detection information for training and its corresponding GT (Ground Truth), and (iii) lea rn ing at least part of one or more parameters of the pedestrian detector to minimize the first losses; configured to execute the instructions to: perform processes of: (I) if at least one test image with each of bounding boxes for testing corresponding to each of pedestrians for testing generated thereon is acquired, generating one or more image patches for testing by cropping each of regions for testing, corresponding to each of the bounding boxes for testing, on the test image, and instructing an adversarial style transformer to generate one or more transformed image patches for testing by converting each of the pedestrians for testing corresponding to each of the image patches for testing into each of transformed pedestrians for testing capable of impeding a detection for testing by the pedestrian detector, and (II) generating at least one transformed test image by replacing each of the regions for testing, corresponding to each of the bounding boxes for testing, on the test image with the transformed image patches for testing, and instructing the pedestrian detector to generate pedestrian detection information for testing by detecting the transformed pedestrians for testing located in the transformed test image, wherein, at the process of (I), the processor resizes at least one of the image patches for testing such that the image patches for testing have a same size, and instructs the adversarial style transformer to output the transformed image patches for testing with the same size, and wherein, at the process of (II), the processor resizes the transformed image patches for testing with the same size such that each of the transformed image patches for testing has its own corresponding original size before the transformation, and then generates the transformed test image. Currently amended
The testing device of Claim 22, wherein, at the process of (II), the processor instructs a discriminator to generate each of pedestrian scores for testing representing each of probabilities of each of the transformed image patches for testing being each of the pedestrians for testing. Original
The testing device of Claim 22, wherein the adversarial style transformer includes an encoder having one or more convolutional layers for applying one or more convolution operations to the image patches for testing, and a decoder having one or more deconvolutional layers for applying one or more deconvolution operations to at least one feature map for testing outputted from the encoder. Original
The testing device of Claim 22, wherein the training image has been acquired from one of (i) the pedestrian detector outputting the pedestrian detection information for testing including the bounding boxes for testing by detecting the pedestrians for testing located in the test image acquired from a surveillance camera, and (ii) a database storing the training image with its own true labels corresponding to the bounding boxes for training. Original
Canceled
Patent
Atlas literature
Patent
US 10,692,002Patent drawings and their descriptions. Click a drawing to enlarge it.
Claims define the patent's legal scope. Independent claims stand alone; dependent claims (nested) narrow them. Click a claim to expand its dependents.
A method for learning a pedestrian detector to be used for robust surveillance based on image analysis, comprising steps of: (a) a learning device, if at least one training image with each of bounding boxes corresponding to each of pedestrians for training generated thereon is acquired, generating one or more image patches by cropping each of regions, corresponding to each of the bounding boxes, on the training image, and instructing an adversarial style transformer to generate one or more transformed image patches by converting each of the pedestrians for training corresponding to each of the image patches into each of transformed pedestrians capable of impeding a detection by the pedestrian detector; and (b) the learning device generating at least one transformed training image by replacing each of the regions, corresponding to each of the bounding boxes, on the training image with the transformed image patches, and learning the pedestrian detector by usin g the transformed ima g e patches in which the pedestrians are more dif fi cult to detect by the pedestrian detector than the one or more image patches from the training image, wherein the step of learning the pedestrian detector includes (i) instructing the pedestrian detector to generate pedestrian detection information for training by detecting the transformed pedestrians located in the transformed training image, (ii) instructing a fi rst loss layer to calculate one or more fi rst losses by referring to each piece of the pedestrian detection information for training and its corresponding GT (Ground Truth), and (iii) learning at least part of one or more parameters of the pedestrian detector to minimize the fi rst losses, wherein, at the step of (a), the learning device resizes at least one of the image patches such that the image patches have a same size, and instructs the adversarial style transformer to output the transformed image patches with the same size, and wherein, at the step of (b), the learning device resizes the transformed image patches with the same size such that each of the transformed image patches has its own corresponding original size before the transformation, and then generates the transformed training image. Currently amended
The method of Claim 1, wherein, at the step of (b), the learning device instructs a discriminator to generate each of pedestrian scores for training representing each of probabilities of each of the transformed image patches being each of the pedestrians for training, and to further learn at least part of one or more parameters of the adversarial style transformer to maximize the pedestrian scores for training and the first losses. Original
The method of Claim 1, wherein the adversarial style transformer includes an encoder having one or more convolutional layers for applying one or more convolution operations to the image patches, and a decoder having one or more deconvolutional layers for applying one or more deconvolution operations to at least one feature map outputted from the encoder. Original
The method of Claim 1, wherein the training image is acquired from one of (i) the pedestrian detector outputting pedestrian detection information for testing including the bounding boxes by detecting pedestrians for testing located in at least one test image acquired from a surveillance camera, and (ii) a database storing the training image with its own true labels corresponding to the bounding boxes. Original
Canceled
A testing method for a pedestrian detector for testing to be used for robust surveillance based on image analysis, comprising steps of (a) a testing device, on condition that (1) a learning device has generated one or more image patches for training by cropping each of regions for training, corresponding to each of the bounding boxes for training, on at least one training image, which includes each of the bounding boxes for training corresponding to each of pedestrians for training generated thereon, and has instructed an adversarial style transformer to generate one or more transformed image patches for training by converting each of the pedestrians for training corresponding to each of the image patches for training into each of transformed pedestrians for training capable of impeding a detection for training by the pedestrian detector, and (2) the learning device has been generated at least one transformed training image by replacing each of the regions for training, corresponding to each of the bounding boxes for training, on the training image with the transformed image patches for training, has instructed the pedestrian detector to generate pedestrian detection information for training by detecting the transformed pedestrians for training located in the transformed training image, has instructed a first loss layer to calculate one or more first losses by referring to each piece of the pedestrian detection information for training and its corresponding GT (Ground Truth), and has learned at least part of one or more parameters of the pedestrian detector to minimize the first losses; if at least one test image with each of bounding boxes for testing corresponding to each of pedestrians for testing generated thereon is acquired, generating one or more image patches for testing by cropping each of regions for testing, corresponding to each of the bounding boxes for testing, on the test image, and instructing an adversarial style transformer to generate one or more transformed image patches for testing by converting each of the pedestrians for testing corresponding to each of the image patches for testing into each of transformed pedestrians for testing capable of impeding a detection for testing by the pedestrian detector; and (b) the testing device generating at least one transformed test image by replacing each of the regions for testing, corresponding to each of the bounding boxes for testing, on the test image with the transformed image patches for testing, and learning the pedestrian detector by using the transformed image patches in which the pedestrians are more difficult to detect by the pedestrian detector than the one or more ima g e patches from the trainin g image, wherein the lea ming the pedestrian detector includes [[and]] instructing the pedestrian detector to generate pedestrian detection information for testing by detecting the transformed pedestrians for testing located in the transformed test image; wherein, at the step of (a), the learning device resizes at least one of the image patches such that the image patches have a same size, and instructs the adversarial style transformer to output the transformed image patches with the same size, and wherein, at the step of (b), the learning device resizes the transformed image patches with the same size such that each of the transformed image patches has its own corresponding original size before the transformation, and then generates the transformed training image. Currently amended
The testing method of Claim 8, wherein, at the step of (b), the testing device instructs a discriminator to generate each of pedestrian scores for testing representing each of probabilities of each of the transformed image patches for testing being each of the pedestrians for testing. Original
The testing method of Claim 8, wherein the adversarial style transformer includes an encoder having one or more convolutional layers for applying one or more convolution operations to the image patches for testing, and a decoder having one or more deconvolutional layers for applying one or more deconvolution operations to at least one feature map for testing outputted from the encoder. Original
The testing method of Claim 8, wherein the training image has been acquired from one of (i) the pedestrian detector outputting the pedestrian detection information for testing including the bounding boxes for testing by detecting the pedestrians for testing located in the test image acquired from a surveillance camera, and (ii) a database storing the training image with its own true labels corresponding to the bounding boxes for training. Original
Canceled
A learning device for learning a pedestrian detector to be used for robust surveillance based on image analysis, comprising: at least one memory that stores instructions; and at least one processor configured to execute the instructions to: perform processes of (I) if at least one training image with each of bounding boxes corresponding to each of pedestrians for training generated thereon is acquired, generating one or more image patches by cropping each of regions, corresponding to each of the bounding boxes, on the training image, and instructing an adversarial style transformer to generate one or more transformed image patches by converting each of the pedestrians for training corresponding to each of the image patches into each of transformed pedestrians capable of impeding a detection by the pedestrian detector, and (II) generating at least one transformed training image by replacing each of the regions, corresponding to each of the bounding boxes, on the training image with the transformed image patches, and learning the pedestrian detector by using the transformed image patches in which the pedestrians are more dif fi cult to detect by the pedestrian detector than the one or more image patches from the training image, wherein the learning the pedestrian detector includes (i) instructing the pedestrian detector to generate pedestrian detection information for training by detecting the transformed pedestrians located in the transformed training image, (ii) instructing a fi rst loss layer to calculate one or more fi rst losses by referring to each piece of the pedestrian detection information for training and its corresponding GT (Ground Truth), and (iii) lea rn ing at least part of one or more parameters of the pedestrian detector to minimize the fi rst losses, wherein, at the process of (I), the learning device resizes at least one of the image patches such that the image patches have a same size, and instructs the adversarial style transformer to output the transformed image patches with the same size, and wherein, at the process of (II), the learning device resizes the transformed image patches with the same size such that each of the transformed image patches has its own corresponding original size before the transformation, and then generates the transformed training image. Currently amended
The learning device of Claim 15, wherein, at the process of (II), the processor instructs a discriminator to generate each of pedestrian scores for training representing each of probabilities of each of the transformed image patches being each of the pedestrians for training, and to further learn at least part of one or more parameters of the adversarial style transformer to maximize the pedestrian scores for training and the first losses. Original
The learning device of Claim 15, wherein the adversarial style transformer includes an encoder having one or more convolutional layers for applying one or more convolution operations to the image patches, and a decoder having one or more deconvolutional layers for applying one or more deconvolution operations to at least one feature map outputted from the encoder. Original
The learning device of Claim 15, wherein the training image is acquired from one of (i) the pedestrian detector outputting pedestrian detection information for testing including the bounding boxes by detecting pedestrians for testing located in at least one test image acquired from a surveillance camera, and (ii) a database storing the training image with its own true labels corresponding to the bounding boxes. Original
Canceled
A testing device for a pedestrian detector for testing to be used for robust surveillance based on image analysis, comprising: at least one memory that stores instructions; and at least one processor, on condition that (1) a learning device has generated one or more image patches for training by cropping each of regions for training, corresponding to each of the bounding boxes for training, on at least one training image, which includes each of the bounding boxes for training corresponding to each of pedestrians for training generated thereon, and has instructed an adversarial style transformer to generate one or more transformed image patches for training by converting each of the pedestrians for training corresponding to each of the image patches for training into each of transformed pedestrians for training capable of impeding a detection for training by the pedestrian detector, and (2) the learning device has [[been]] generated at least one transformed training image by replacing each of the regions for training, corresponding to each of the bounding boxes for training, on the training image with the transformed image patches for training, and has learned the pedestrian detector by usin g the transformed image patches in which the pedestrians are more difficult to detect by the pedestrian detector than the one or more image patches from the training image, wherein the learning of the pedestrian detector included (i) instructing the pedestrian detector to generate pedestrian detection information for training by detecting the transformed pedestrians for training located in the transformed training image, (ii) instructing a first loss layer to calculate one or more first losses by referring to each piece of the pedestrian detection information for training and its corresponding GT (Ground Truth), and (iii) lea rn ing at least part of one or more parameters of the pedestrian detector to minimize the first losses; configured to execute the instructions to: perform processes of: (I) if at least one test image with each of bounding boxes for testing corresponding to each of pedestrians for testing generated thereon is acquired, generating one or more image patches for testing by cropping each of regions for testing, corresponding to each of the bounding boxes for testing, on the test image, and instructing an adversarial style transformer to generate one or more transformed image patches for testing by converting each of the pedestrians for testing corresponding to each of the image patches for testing into each of transformed pedestrians for testing capable of impeding a detection for testing by the pedestrian detector, and (II) generating at least one transformed test image by replacing each of the regions for testing, corresponding to each of the bounding boxes for testing, on the test image with the transformed image patches for testing, and instructing the pedestrian detector to generate pedestrian detection information for testing by detecting the transformed pedestrians for testing located in the transformed test image, wherein, at the process of (I), the processor resizes at least one of the image patches for testing such that the image patches for testing have a same size, and instructs the adversarial style transformer to output the transformed image patches for testing with the same size, and wherein, at the process of (II), the processor resizes the transformed image patches for testing with the same size such that each of the transformed image patches for testing has its own corresponding original size before the transformation, and then generates the transformed test image. Currently amended
The testing device of Claim 22, wherein, at the process of (II), the processor instructs a discriminator to generate each of pedestrian scores for testing representing each of probabilities of each of the transformed image patches for testing being each of the pedestrians for testing. Original
The testing device of Claim 22, wherein the adversarial style transformer includes an encoder having one or more convolutional layers for applying one or more convolution operations to the image patches for testing, and a decoder having one or more deconvolutional layers for applying one or more deconvolution operations to at least one feature map for testing outputted from the encoder. Original
The testing device of Claim 22, wherein the training image has been acquired from one of (i) the pedestrian detector outputting the pedestrian detection information for testing including the bounding boxes for testing by detecting the pedestrians for testing located in the test image acquired from a surveillance camera, and (ii) a database storing the training image with its own true labels corresponding to the bounding boxes for training. Original
Canceled
Patent
Atlas literature
Patent
US 10,692,002Patent drawings and their descriptions. Click a drawing to enlarge it.
Claims define the patent's legal scope. Independent claims stand alone; dependent claims (nested) narrow them. Click a claim to expand its dependents.
A method for learning a pedestrian detector to be used for robust surveillance based on image analysis, comprising steps of: (a) a learning device, if at least one training image with each of bounding boxes corresponding to each of pedestrians for training generated thereon is acquired, generating one or more image patches by cropping each of regions, corresponding to each of the bounding boxes, on the training image, and instructing an adversarial style transformer to generate one or more transformed image patches by converting each of the pedestrians for training corresponding to each of the image patches into each of transformed pedestrians capable of impeding a detection by the pedestrian detector; and (b) the learning device generating at least one transformed training image by replacing each of the regions, corresponding to each of the bounding boxes, on the training image with the transformed image patches, and learning the pedestrian detector by usin g the transformed ima g e patches in which the pedestrians are more dif fi cult to detect by the pedestrian detector than the one or more image patches from the training image, wherein the step of learning the pedestrian detector includes (i) instructing the pedestrian detector to generate pedestrian detection information for training by detecting the transformed pedestrians located in the transformed training image, (ii) instructing a fi rst loss layer to calculate one or more fi rst losses by referring to each piece of the pedestrian detection information for training and its corresponding GT (Ground Truth), and (iii) learning at least part of one or more parameters of the pedestrian detector to minimize the fi rst losses, wherein, at the step of (a), the learning device resizes at least one of the image patches such that the image patches have a same size, and instructs the adversarial style transformer to output the transformed image patches with the same size, and wherein, at the step of (b), the learning device resizes the transformed image patches with the same size such that each of the transformed image patches has its own corresponding original size before the transformation, and then generates the transformed training image. Currently amended
The method of Claim 1, wherein, at the step of (b), the learning device instructs a discriminator to generate each of pedestrian scores for training representing each of probabilities of each of the transformed image patches being each of the pedestrians for training, and to further learn at least part of one or more parameters of the adversarial style transformer to maximize the pedestrian scores for training and the first losses. Original
The method of Claim 1, wherein the adversarial style transformer includes an encoder having one or more convolutional layers for applying one or more convolution operations to the image patches, and a decoder having one or more deconvolutional layers for applying one or more deconvolution operations to at least one feature map outputted from the encoder. Original
The method of Claim 1, wherein the training image is acquired from one of (i) the pedestrian detector outputting pedestrian detection information for testing including the bounding boxes by detecting pedestrians for testing located in at least one test image acquired from a surveillance camera, and (ii) a database storing the training image with its own true labels corresponding to the bounding boxes. Original
Canceled
A testing method for a pedestrian detector for testing to be used for robust surveillance based on image analysis, comprising steps of (a) a testing device, on condition that (1) a learning device has generated one or more image patches for training by cropping each of regions for training, corresponding to each of the bounding boxes for training, on at least one training image, which includes each of the bounding boxes for training corresponding to each of pedestrians for training generated thereon, and has instructed an adversarial style transformer to generate one or more transformed image patches for training by converting each of the pedestrians for training corresponding to each of the image patches for training into each of transformed pedestrians for training capable of impeding a detection for training by the pedestrian detector, and (2) the learning device has been generated at least one transformed training image by replacing each of the regions for training, corresponding to each of the bounding boxes for training, on the training image with the transformed image patches for training, has instructed the pedestrian detector to generate pedestrian detection information for training by detecting the transformed pedestrians for training located in the transformed training image, has instructed a first loss layer to calculate one or more first losses by referring to each piece of the pedestrian detection information for training and its corresponding GT (Ground Truth), and has learned at least part of one or more parameters of the pedestrian detector to minimize the first losses; if at least one test image with each of bounding boxes for testing corresponding to each of pedestrians for testing generated thereon is acquired, generating one or more image patches for testing by cropping each of regions for testing, corresponding to each of the bounding boxes for testing, on the test image, and instructing an adversarial style transformer to generate one or more transformed image patches for testing by converting each of the pedestrians for testing corresponding to each of the image patches for testing into each of transformed pedestrians for testing capable of impeding a detection for testing by the pedestrian detector; and (b) the testing device generating at least one transformed test image by replacing each of the regions for testing, corresponding to each of the bounding boxes for testing, on the test image with the transformed image patches for testing, and learning the pedestrian detector by using the transformed image patches in which the pedestrians are more difficult to detect by the pedestrian detector than the one or more ima g e patches from the trainin g image, wherein the lea ming the pedestrian detector includes [[and]] instructing the pedestrian detector to generate pedestrian detection information for testing by detecting the transformed pedestrians for testing located in the transformed test image; wherein, at the step of (a), the learning device resizes at least one of the image patches such that the image patches have a same size, and instructs the adversarial style transformer to output the transformed image patches with the same size, and wherein, at the step of (b), the learning device resizes the transformed image patches with the same size such that each of the transformed image patches has its own corresponding original size before the transformation, and then generates the transformed training image. Currently amended
The testing method of Claim 8, wherein, at the step of (b), the testing device instructs a discriminator to generate each of pedestrian scores for testing representing each of probabilities of each of the transformed image patches for testing being each of the pedestrians for testing. Original
The testing method of Claim 8, wherein the adversarial style transformer includes an encoder having one or more convolutional layers for applying one or more convolution operations to the image patches for testing, and a decoder having one or more deconvolutional layers for applying one or more deconvolution operations to at least one feature map for testing outputted from the encoder. Original
The testing method of Claim 8, wherein the training image has been acquired from one of (i) the pedestrian detector outputting the pedestrian detection information for testing including the bounding boxes for testing by detecting the pedestrians for testing located in the test image acquired from a surveillance camera, and (ii) a database storing the training image with its own true labels corresponding to the bounding boxes for training. Original
Canceled
A learning device for learning a pedestrian detector to be used for robust surveillance based on image analysis, comprising: at least one memory that stores instructions; and at least one processor configured to execute the instructions to: perform processes of (I) if at least one training image with each of bounding boxes corresponding to each of pedestrians for training generated thereon is acquired, generating one or more image patches by cropping each of regions, corresponding to each of the bounding boxes, on the training image, and instructing an adversarial style transformer to generate one or more transformed image patches by converting each of the pedestrians for training corresponding to each of the image patches into each of transformed pedestrians capable of impeding a detection by the pedestrian detector, and (II) generating at least one transformed training image by replacing each of the regions, corresponding to each of the bounding boxes, on the training image with the transformed image patches, and learning the pedestrian detector by using the transformed image patches in which the pedestrians are more dif fi cult to detect by the pedestrian detector than the one or more image patches from the training image, wherein the learning the pedestrian detector includes (i) instructing the pedestrian detector to generate pedestrian detection information for training by detecting the transformed pedestrians located in the transformed training image, (ii) instructing a fi rst loss layer to calculate one or more fi rst losses by referring to each piece of the pedestrian detection information for training and its corresponding GT (Ground Truth), and (iii) lea rn ing at least part of one or more parameters of the pedestrian detector to minimize the fi rst losses, wherein, at the process of (I), the learning device resizes at least one of the image patches such that the image patches have a same size, and instructs the adversarial style transformer to output the transformed image patches with the same size, and wherein, at the process of (II), the learning device resizes the transformed image patches with the same size such that each of the transformed image patches has its own corresponding original size before the transformation, and then generates the transformed training image. Currently amended
The learning device of Claim 15, wherein, at the process of (II), the processor instructs a discriminator to generate each of pedestrian scores for training representing each of probabilities of each of the transformed image patches being each of the pedestrians for training, and to further learn at least part of one or more parameters of the adversarial style transformer to maximize the pedestrian scores for training and the first losses. Original
The learning device of Claim 15, wherein the adversarial style transformer includes an encoder having one or more convolutional layers for applying one or more convolution operations to the image patches, and a decoder having one or more deconvolutional layers for applying one or more deconvolution operations to at least one feature map outputted from the encoder. Original
The learning device of Claim 15, wherein the training image is acquired from one of (i) the pedestrian detector outputting pedestrian detection information for testing including the bounding boxes by detecting pedestrians for testing located in at least one test image acquired from a surveillance camera, and (ii) a database storing the training image with its own true labels corresponding to the bounding boxes. Original
Canceled
A testing device for a pedestrian detector for testing to be used for robust surveillance based on image analysis, comprising: at least one memory that stores instructions; and at least one processor, on condition that (1) a learning device has generated one or more image patches for training by cropping each of regions for training, corresponding to each of the bounding boxes for training, on at least one training image, which includes each of the bounding boxes for training corresponding to each of pedestrians for training generated thereon, and has instructed an adversarial style transformer to generate one or more transformed image patches for training by converting each of the pedestrians for training corresponding to each of the image patches for training into each of transformed pedestrians for training capable of impeding a detection for training by the pedestrian detector, and (2) the learning device has [[been]] generated at least one transformed training image by replacing each of the regions for training, corresponding to each of the bounding boxes for training, on the training image with the transformed image patches for training, and has learned the pedestrian detector by usin g the transformed image patches in which the pedestrians are more difficult to detect by the pedestrian detector than the one or more image patches from the training image, wherein the learning of the pedestrian detector included (i) instructing the pedestrian detector to generate pedestrian detection information for training by detecting the transformed pedestrians for training located in the transformed training image, (ii) instructing a first loss layer to calculate one or more first losses by referring to each piece of the pedestrian detection information for training and its corresponding GT (Ground Truth), and (iii) lea rn ing at least part of one or more parameters of the pedestrian detector to minimize the first losses; configured to execute the instructions to: perform processes of: (I) if at least one test image with each of bounding boxes for testing corresponding to each of pedestrians for testing generated thereon is acquired, generating one or more image patches for testing by cropping each of regions for testing, corresponding to each of the bounding boxes for testing, on the test image, and instructing an adversarial style transformer to generate one or more transformed image patches for testing by converting each of the pedestrians for testing corresponding to each of the image patches for testing into each of transformed pedestrians for testing capable of impeding a detection for testing by the pedestrian detector, and (II) generating at least one transformed test image by replacing each of the regions for testing, corresponding to each of the bounding boxes for testing, on the test image with the transformed image patches for testing, and instructing the pedestrian detector to generate pedestrian detection information for testing by detecting the transformed pedestrians for testing located in the transformed test image, wherein, at the process of (I), the processor resizes at least one of the image patches for testing such that the image patches for testing have a same size, and instructs the adversarial style transformer to output the transformed image patches for testing with the same size, and wherein, at the process of (II), the processor resizes the transformed image patches for testing with the same size such that each of the transformed image patches for testing has its own corresponding original size before the transformation, and then generates the transformed test image. Currently amended
The testing device of Claim 22, wherein, at the process of (II), the processor instructs a discriminator to generate each of pedestrian scores for testing representing each of probabilities of each of the transformed image patches for testing being each of the pedestrians for testing. Original
The testing device of Claim 22, wherein the adversarial style transformer includes an encoder having one or more convolutional layers for applying one or more convolution operations to the image patches for testing, and a decoder having one or more deconvolutional layers for applying one or more deconvolution operations to at least one feature map for testing outputted from the encoder. Original
The testing device of Claim 22, wherein the training image has been acquired from one of (i) the pedestrian detector outputting the pedestrian detection information for testing including the bounding boxes for testing by detecting the pedestrians for testing located in the test image acquired from a surveillance camera, and (ii) a database storing the training image with its own true labels corresponding to the bounding boxes for training. Original
Canceled
