Superb Ai Co., Ltd.

Kye Hyeon Kim

Hyundong Lee

The present disclosure provides a method for training a custom model based on a pre-trained base model. The custom model is configured to include a base model having at least one pre-trained feature extraction block, at least one adaptation block, at least one residual unit, and a task head block, and update at least part of the task head block, the at least one residual unit and the at least one adaptation block through a backpropagation using a task loss, while not updating parameters of the base model. Further, at least one residual feature (generated through the at least one residual unit) and at least one converted feature (generated by converting the at least one intermediate feature through the at least one adaptation block) are fused to thereby allow a learning device to achieve an advanced inference.

Method for training custom model based on pre-trained base model and learning device using the same

A method for selecting training data for training a deep learning model is provided. The method includes steps of: (a) obtaining one or more individual attributes each of which corresponds to each of a plurality of training data included in total training data, and generating a bipartite graph by matching each of the plurality of training data included in the total training data with the individual attributes; and (b) selecting n training data among the total training data, by referring to the bipartite graph, wherein the n is a target number of the training data to be used for training the deep learning model, and wherein the training data selecting device selects the n training data to be used for training the deep learning model such that each cardinal number of each of the individual attributes matched with the n training data is within a predetermined deviation threshold.

Method for selecting training data to train a deep learning model and training data selecting device using the same

There is provided a learning method and a learning device for performing transfer learning on an object detector that has been trained to detect first object classes such that the object detector is able to detect second object classes. Further, a testing method and a testing device are provided to allow at least part of the first object classes and the second object classes to be detected by using the object detector having been trained through the transfer learning. Accordingly, a detection performance can be improved for the second object classes that cannot be detected through training data set corresponding to the first object classes.

Learning method and learning device for performing transfer learning on an object detector that has been trained to detect first object classes such that the object detector is able to detect second object classes, and testing method and testing device using the same

A method for generating an object detector based on deep learning capable of detecting an extended object class is provided. The method is related to generating the object detector based on the deep learning capable of detecting the extended object class to thereby allow both an object class having been trained and additional object class to be detected. According to the method, it is possible to generate the training data set necessary for training an object detector capable of detecting the extended object class at a low cost in a short time and further it is possible to generate the object detector capable of detecting the extended object class at a low cost in a short time.

Method for generating object detector based on deep learning capable of detecting extended object class and detector generation device using the same

A method for tracking an object in a low frame rate video is provided. Matching processes are performed between consecutive frames by using conversion feature maps acquired by converting each of features on feature maps of the consecutive frames into feature descriptors including each corresponding feature information and each corresponding location information, thereby allowing object tracking regardless of whether time interval per frame is long or short. The object tracking is performed by matching feature descriptors on a plurality of pyramid feature maps on an entire area of a next frame and feature descriptors on a plurality of cropped feature maps generated by cropping object areas extracted on a current frame, thereby allowing not only quick matching between the cropped areas and the entire area but also the increased accuracy due to no limitation of the feature searching area.

Growing community of inventors

Suwon-si, South Korea

Kye Hyeon Kim