-
公开(公告)号：US11315021B2
公开(公告)日：2022-04-26
申请号：US16259389
申请日：2019-01-28
申请人： Stradvision, Inc.
发明人： Kye-Hyeon Kim , Yongjoong Kim , Insu Kim , Hak-Kyoung Kim , Woonhyun Nam , SukHoon Boo , Myungchul Sung , Donghun Yeo , Wooju Ryu , Taewoong Jang , Kyungjoong Jeong , Hongmo Je , Hojin Cho
摘要： A method for on-device continual learning of a neural network which analyzes input data is provided to be used for smartphones, drones, vessels, or a military purpose. The method includes steps of: a learning device, (a) sampling new data to have a preset first volume, instructing an original data generator network, which has been learned, to repeat outputting synthetic previous data corresponding to a k-dimension random vector and previous data having been used for learning the original data generator network, such that the synthetic previous data has a second volume, and generating a batch for a current-learning; and (b) instructing the neural network to generate output information corresponding to the batch. The method can be performed by generative adversarial networks (GANs), online learning, and the like. Also, the present disclosure has effects of saving resources such as storage, preventing catastrophic forgetting, and securing privacy.
-
公开(公告)号：US20200234135A1
公开(公告)日：2020-07-23
申请号：US16255109
申请日：2019-01-23
申请人： Stradvision, Inc.
发明人： Kye-Hyeon Kim , Yongjoong Kim , Insu Kim , Hak-Kyoung Kim , Woonhyun Nam , SukHoon Boo , Myungchul Sung , Donghun Yeo , Wooju Ryu , Taewoong Jang , Kyungjoong Jeong , Hongmo Je , Hojin Cho
摘要： A method for learning parameters of a CNN using a 1×K convolution operation or a K×1 convolution operation is provided to be used for hardware optimization which satisfies KPI. The method includes steps of: a learning device (a) instructing a reshaping layer to two-dimensionally concatenate features in each group comprised of corresponding K channels of a training image or its processed feature map, to thereby generate a reshaped feature map, and instructing a subsequent convolutional layer to apply the 1×K or the K×1 convolution operation to the reshaped feature map, to thereby generate an adjusted feature map; and (b) instructing an output layer to refer to features on the adjusted feature map or its processed feature map, and instructing a loss layer to calculate losses by referring to an output from the output layer and its corresponding GT.
-
公开(公告)号：US10699192B1
公开(公告)日：2020-06-30
申请号：US16262985
申请日：2019-01-31
申请人： Stradvision, Inc.
发明人： Kye-Hyeon Kim , Yongjoong Kim , Insu Kim , Hak-Kyoung Kim , Woonhyun Nam , SukHoon Boo , Myungchul Sung , Donghun Yeo , Wooju Ryu , Taewoong Jang , Kyungjoong Jeong , Hongmo Je , Hojin Cho
摘要： A method for optimizing a hyperparameter of an auto-labeling device performing auto-labeling and auto-evaluating of a training image to be used for learning a neural network is provided for computation reduction and achieving high precision. The method includes steps of: an optimizing device, (a) instructing the auto-labeling device to generate an original image with its auto label and a validation image with its true and auto label, to assort the original image with its auto label into an easy-original and a difficult-original images, and to assort the validation image with its own true and auto labels into an easy-validation and a difficult-validation images; and (b) calculating a current reliability of the auto-labeling device, generating a sample hyperparameter set, calculating a sample reliability of the auto-labeling device, and optimizing the preset hyperparameter set. This method can be performed by a reinforcement learning with policy gradient algorithms.
-
公开(公告)号：US10482584B1
公开(公告)日：2019-11-19
申请号：US16262996
申请日：2019-01-31
申请人： Stradvision, Inc.
发明人： Kye-Hyeon Kim , Yongjoong Kim , Insu Kim , Hak-Kyoung Kim , Woonhyun Nam , SukHoon Boo , Myungchul Sung , Donghun Yeo , Wooju Ryu , Taewoong Jang , Kyungjoong Jeong , Hongmo Je , Hojin Cho
摘要： A method for detecting jittering in videos generated by a shaken camera to remove the jittering on the videos using neural networks is provided for fault tolerance and fluctuation robustness in extreme situations. The method includes steps of: a computing device, generating each of t-th masks corresponding to each of objects in a t-th image; generating each of t-th object motion vectors of each of object pixels, included in the t-th image by applying at least one 2-nd neural network operation to each of the t-th masks, each of t-th cropped images, each of (t−1)-th masks, and each of (t−1)-th cropped images; and generating each of t-th jittering vectors corresponding to each of reference pixels among pixels in the t-th image by referring to each of the t-th object motion vectors. Thus, the method is used for video stabilization, object tracking with high precision, behavior estimation, motion decomposition, etc.
-
公开(公告)号：US10474543B1
公开(公告)日：2019-11-12
申请号：US16258850
申请日：2019-01-28
申请人： Stradvision, Inc.
发明人： Kye-Hyeon Kim , Yongjoong Kim , Insu Kim , Hak-Kyoung Kim , Woonhyun Nam , SukHoon Boo , Myungchul Sung , Donghun Yeo , Wooju Ryu , Taewoong Jang , Kyungjoong Jeong , Hongmo Je , Hojin Cho
摘要： A method for economizing computing resources and verifying an integrity of parameters of a neural network by inserting test pattern into a background area of an input image is provided for fault tolerance, fluctuation robustness in extreme situations, functional safety on the neural network, and an annotation cost reduction. The method includes: a computing device (a) generating t-th background prediction information of a t-th image by referring to information on each of a (t−2)-th image and a (t−1)-th image; (b) inserting the test pattern into the t-th image by referring to the t-th background prediction information, to thereby generate an input for verification; (c) generating an output for verification from the input for verification; and (d) determining the integrity of the neural network by referring to the output for verification and an output for reference. According to the method, a data compression and a computation reduction are achieved.
-
公开(公告)号：US10325352B1
公开(公告)日：2019-06-18
申请号：US16255197
申请日：2019-01-23
申请人： Stradvision, Inc.
发明人： Kye-Hyeon Kim , Yongjoong Kim , Insu Kim , Hak-Kyoung Kim , Woonhyun Nam , SukHoon Boo , Myungchul Sung , Donghun Yeo , Wooju Ryu , Taewoong Jang , Kyungjoong Jeong , Hongmo Je , Hojin Cho
摘要： There is provided a method for transforming convolutional layers of a CNN including m convolutional blocks to optimize CNN parameter quantization to be used for mobile devices, compact networks, and the like with high precision via hardware optimization. The method includes steps of: a computing device (a) generating k-th quantization loss values by referring to k-th initial weights of a k-th initial convolutional layer included in a k-th convolutional block, a (k−1)-th feature map outputted from the (k−1)-th convolutional block, and each of k-th scaling parameters; (b) determining each of k-th optimized scaling parameters by referring to the k-th quantization loss values; (c) generating a k-th scaling layer and a k-th inverse scaling layer by referring to the k-th optimized scaling parameters; and (d) transforming the k-th initial convolutional layer into a k-th integrated convolutional layer by using the k-th scaling layer and the (k−1)-th inverse scaling layer.
-
公开(公告)号：US10229346B1
公开(公告)日：2019-03-12
申请号：US16120664
申请日：2018-09-04
申请人： Stradvision, Inc.
发明人： Kye-Hyeon Kim , Yongjoong Kim , Insu Kim , Hak-Kyoung Kim , Woonhyun Nam , SukHoon Boo , Myungchul Sung , Donghun Yeo , Wooju Ryu , Taewoong Jang , Kyungjoong Jeong , Hongmo Je , Hojin Cho
摘要： A learning method for detecting a specific object based on convolutional neural network (CNN) is provided. The learning method includes steps of: (a) a learning device, if an input image is obtained, performing (i) a process of applying one or more convolution operations to the input image to thereby obtain at least one specific feature map and (ii) a process of obtaining an edge image by extracting at least one edge part from the input image, and obtaining at least one guide map including information on at least one specific edge part having a specific shape similar to that of the specific object from the obtained edge image; and (b) the learning device reflecting the guide map on the specific feature map to thereby obtain a segmentation result for detecting the specific object in the input image.
-
公开(公告)号：US10223614B1
公开(公告)日：2019-03-05
申请号：US16121084
申请日：2018-09-04
申请人： Stradvision, Inc.
发明人： Kye-Hyeon Kim , Yongjoong Kim , Insu Kim , Hak-Kyoung Kim , Woonhyun Nam , SukHoon Boo , Myungchul Sung , Donghun Yeo , Wooju Ryu , Taewoong Jang , Kyungjoong Jeong , Hongmo Je , Hojin Cho
摘要： A learning method for detecting at least one lane based on a convolutional neural network (CNN) is provided. The learning method includes steps of: (a) a learning device obtaining encoded feature maps, and information on lane candidate pixels in a input image; (b) the learning device, classifying a first parts of the lane candidate pixels, whose probability scores are not smaller than a predetermined threshold, as strong line pixels, and classifying the second parts of the lane candidate pixels, whose probability scores are less than the threshold but not less than another predetermined threshold, as weak lines pixels; and (c) the learning device, if distances between the weak line pixels and the strong line pixels are less than a predetermined distance, classifying the weak line pixels as pixels of additional strong lines, and determining that the pixels of the strong line and the additional correspond to pixels of the lane.
-
9.发明申请公开(公告)号：US20200090047A1
公开(公告)日：2020-03-19
申请号：US16132479
申请日：2018-09-17
申请人： Stradvision, Inc.
发明人： Kye-Hyeon Kim , Yongjoong Kim , Insu Kim , Hak-Kyoung Kim , Woonhyun Nam , SukHoon Boo , Myungchul Sung , Donghun Yeo , Wooju Ryu , Taewoong Jang , Kyungjoong Jeong , Hongmo Je , Hojin Cho
摘要： A learning method for a CNN (Convolutional Neural Network) capable of encoding at least one training image with multiple feeding layers, wherein the CNN includes a 1st to an n-th convolutional layers, which respectively generate a 1st to an n-th main feature maps by applying convolution operations to the training image, and a 1st to an h-th feeding layers respectively corresponding to h convolutional layers (1≤h≤(n-1)) is provided. The learning method includes steps of: a learning device instructing the convolutional layers to generate the 1st to the n-th main feature maps, wherein the learning device instructs a k-th convolutional layer to acquire a (k−1)-th main feature map and an m-th sub feature map, and to generate a k-th main feature map by applying the convolution operations to the (k−1)-th integrated feature map generated by integrating the (k−1)-th main feature map and the m-th sub feature map.
-
公开(公告)号：US10551846B1
公开(公告)日：2020-02-04
申请号：US16257993
申请日：2019-01-25
申请人： Stradvision, Inc.
发明人： Kye-Hyeon Kim , Yongjoong Kim , Insu Kim , Hak-Kyoung Kim , Woonhyun Nam , SukHoon Boo , Myungchul Sung , Donghun Yeo , Wooju Ryu , Taewoong Jang , Kyungjoong Jeong , Hongmo Je , Hojin Cho
摘要： A learning method for improving segmentation performance to be used for detecting road user events including pedestrian events and vehicle events using double embedding configuration in a multi-camera system is provided. The learning method includes steps of: a learning device instructing similarity convolutional layer to generate similarity embedding feature by applying similarity convolution operations to a feature outputted from a neural network; instructing similarity loss layer to output a similarity loss by referring to a similarity between two points sampled from the similarity embedding feature, and its corresponding GT label image; instructing distance convolutional layer to generate distance embedding feature by applying distance convolution operations to the similarity embedding feature; instructing distance loss layer to output a distance loss for increasing inter-class differences among mean values of instance classes and decreasing intra-class variance values of the instance classes; backpropagating at least one of the similarity loss and the distance loss.
-
-
-
-
-
-
-
-
-
搜索结果
82 条记录 (耗时 0.02 秒)
