Improvement of Object Detection Training Method
- 提供方法
- 本サイト上にてダウンロード・閲覧可
- 形態
- 価格
- 一般価格(税込):¥1,100 会員価格(税込):¥880
- 文献番号
- 20219002
- 文献・情報種別
- その他の国際会議
- 掲載ページ
- 1-1(Total 1 p)
- 発行年月
- 2021年 9月
- 出版社
- (公社)自動車技術会
- 言語
- 英語
書誌事項
著者(英) | 1) Chi-Shen Chao, 2) Mohd Hafiz Hilman Mohammad Sofian, 3) Toshio Ito |
---|---|
勤務先(英) | 1) Shibaura Institute of Technology, 2) Shibaura Institute of Technology, 3) Shibaura Institute of Technology |
抄録(英) | One-stage object detection shows high performance in the field of real-time object detection. Thanks to CNN, we can start to build a high accuracy object detection model. However, due to the working principle of object detection, the quality of the input image and the annotation will affect model accuracy massively. During the training process, the missed annotation in single image might confuse our model. In this paper, we designed two patterns of training method. The first pattern is making multiple annotations in single image. In the second pattern, we only annotate one object in single image and ignore others existed objects in the same image. Finally, we compared the results and check how this problem will affect the final accuracy. 翻訳 |