An Analysis on Samples Selections with Various Confident Levels for Vanilla Pseudo-labeling

• Main Authors: Bor-Hon Tu, 杜柏宏
• Other Authors: Kao-Shing Hwang
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/4w3264

碩士 === 國立中山大學 === 電機工程學系研究所 === 107 === In recent years, deep learning become a very popular research topic. Especially, neural networks have made a successful application in target detection. Its accuracy rate has been catching up with humans or better than humans. The algorithm chosen for this paper is YOLO because of its special feature, only using one convolutional neural network to finish image tracking. The prediction performance of convolutional neural network is not only related to the architecture of algorithm design, but also influenced by the data set for training. More importantly, the training set required for image tracking is not only pictures, but also the coordinate position of the target. Traditionally, this type of training set is artificially produced. but in this thesis, the predictions made on other online videos back are directly put into the training set to augment the training materials and retrain them, then collect and re-organize the forecast data of their new weights. Hopefully, the YOLO development of this thesis results in worth reference for other application.