Summary
Smart Info-Media Systems in Asia
2019
Session Number:SS1
Session:
Number:SS1-2
A Quick Data Generation Method for Training Object Detection Algorithms in Home Environments
Yuma Yoshimoto, Muhammad Farhan Mustafa, Wan Zuha Wan Hasan, Hakaru Tamukoh,
pp.7-10
Publication Date:2019/9/4
Online ISSN:2188-5079
DOI:10.34385/proc.57.SS1-2
PDF download (8.5MB)
Summary:
Deep neural networks are the mainstream of object detection algorithms. The required data are scene images and annotation data for training. Here we propose a new model for training object detection algorithms. Data in our method are quickly generated by a four-step procedure: (1) videos acquisition of objects, (2) saving of video frames as scene images, (3) generation of annotation data from the detection results of You Only Look Once 9000 (YOLOv2), which inputs scene images, and (4) data augmentation. In a comparison experiment, the proposed method generated data 10 times faster than conventional methods. We then trained YOLOv2 on the data generated by the proposed method, and evaluated the effectiveness of the proposed method. The training increased the Intersection-over-Union measure of YOLOv2, confirming the effectiveness of training by the proposed method.