The present disclosure is directed to systems and methods for generating synthetic training data using augmented reality (AR) techniques. For example, images of a scene can be used to generate a three-dimensional mapping of the scene. The three-dimensional mapping may be associated with the images to indicate locations for positioning a virtual object. Using an AR rendering engine, implementations can generate an and orientation. The augmented image can then be stored in a machine learning dataset and associated with a label based on aspects of the virtual object.

An analysis device includes: a memory; and a processor coupled to the memory and configured to: execute a first learning process on a generative model for images such that the images that bring a recognition result of an image recognition process into a preassigned state are generated; execute a second learning process on the generative model on which the first learning process which has been executed such that recognition accuracy of the images generated by the generative model on which the first learning process has been executed matches desired recognition accuracy; acquire information on back-error propagation calculated by executing the image recognition process, for the images with the desired recognition accuracy generated by executing the second learning process; and generate evaluation information that indicates image parts that cause over-detection at the desired recognition accuracy, based on the acquired information on the back-error propagation.

There are provided a method for segmenting a sign language video by gloss to recognize a sign language sentence, and a method for training. According to an embodiment, a sign language video segmentation method receives an input of a sign language sentence video, and segments the inputted sign language sentence video by gloss. Accordingly, there is suggested a method for segmenting a sign language sentence video by gloss, analyzing various gloss sequences from the linguistic perspective, understanding meanings robustly in spite of various changes in sentences, and translating sign language into appropriate Korean sentences.

A system for updating a training dataset of an item identification model determines that an item is not included in a training dataset. In response to determining that the item is not included in the training dataset, the system obtains an identifier of the item. The system detects a triggering event at a platform, where the triggering event corresponds to a user placing the item on a platform. The system captures images of the item. The system extracts a set of features associated with the item from the images. The system associates the item to the identifier and the set of features. The system adds a new entry to the training dataset, where the new entry represents the item labeled with the identifier and the set of features.

A system for identifying items based on aggregated metadata obtains images of an item. The system extracts a set of features from images of the item. The system identifies a first value of a first feature associated with a first image of the item. The system identifies a second value of the first feature associated with a second image of the item. The system aggregates the first value and the second value. The system associates the item to the aggregated first value and the second value, where the aggregated first value and the second value represent the first feature of the item. The system adds a new entry for each image of the item to a training dataset associated with an item identification model.

Systems and methods for bird's eye view (BEV) segmentation are provided. In one embodiment, a method includes receiving an input image from an image sensor on an agent. The input image is a perspective space image defined relative to the position and viewing direction of the agent. The method includes extracting features from the input image. The method includes estimating a depth map that includes depth values for pixels of the plurality of pixels of the input image. The method includes generating a 3D point map including points corresponding to the pixels of the input image. The method includes generating a voxel grid by voxelizing the 3D point map into a plurality voxels. The method includes generating a feature map by extracting feature vectors for pixels based on the points included in the voxels of the plurality of voxels and generating a BEV segmentation based on the feature map.

The purpose of the present invention is to provide a technology for updating a recognition model so that even if there were errors in recognition of an unknown scene or the like, the scene can be recognized quickly. The present invention is provided with a data analysis unit 11 that, on the basis of data from an outside recognition unit 32 provided to a vehicle, acquires from among previously stored recognition models a model approximate to a recognition model recognized by the outside recognition unit 32, and that reproduces the acquired model in the form of computer graphics images. The data analysis unit 11 is provided with: a difference extraction unit 114 that compares the reproduced computer graphics images and data from the outside recognition unit 32 and extracts a difference therebetween; an object recognition unit 116 that recognizes an object relating to the difference extracted by the difference extraction unit 114; and a scene reconfiguration unit 117 that creates computer graphics images having the object recognized by the recognition unit 116 reflected therein.

A neural network model training method and an apparatus for complex characteristic classification and common localization are proposed. In the method, a neural network model includes: a convolution layer for performing a convolution operation on an input image by using a convolution filter; a pooling layer for performing pooling on an output of the convolution layer; and class-specific fully connected layers respectively corresponding to classes into which complex characteristics are classified and outputting values obtained by multiplying an output of the pooling layer by class-specific weights (w.sub.fc(T.sub.t)). The method includes: (a) inputting the input image to the convolution layer; (b) calculating class-specific observation maps for respective classes on the basis of the output of the convolution layer; (c) calculating an observation loss (L.sub.obs) common to the classes on the basis of the class-specific observation maps; and (d) back-propagating a loss based on the observation loss to the neural network model.

A method comprising: receiving a set of images associated with each of a plurality of plants in a plantation; estimating, with respect to each of the plants, based, at least in part, on the set of images associated with the plant, the following data: (i) a count of fruits detected in the plant, and (ii) one or more features associated with the plant; at a training stage, training a machine learning model on a training set comprising, with respect to a subset of the plurality of plants: (iii) the data, and (iv) labels indicating an actual a number of fruits in each of the plants in the subset; and at an inference stage, applying the trained machine learning model to the data associated with the rest of the plurality of plants, to predict a number of fruits in each of the rest of the plurality of plants.

A model generation apparatus according to one or more embodiments executes operations, with respect to each of learning data sets. The operations includes training a second estimator so that an estimation result obtained from a second estimator conforms to second correct answer data; training a coder so that an estimation result obtained from the second estimator does not conform to the second correct answer data; and training the coder and the first estimator so that an estimation result obtained from a first estimator conforms to first correct answer data. The model generation apparatus executes operation of the training the second estimator and the training the coder alternately and repeatedly.