物体間相互損失を用いた選択的インスタンスセグメンテーション

カテゴリ: 論文誌(論文単位)

グループ名: 【C】電子・情報・システム部門

発行日: 2023/12/01

タイトル(英語): Selective Instance Segmentation Using Mutual Loss Between Products

著者名: 日色　紀貴（名城大学 理工学部），田崎　豪（名城大学 理工学部）

著者名(英語): Noritaka Hiiro (Faculty of Science and Technology, Meijo University), Tsuyoshi Tasaki (Faculty of Science and Technology, Meijo University)

キーワード: インスタンスセグメンテーション，把持，深層学習　　instance segmentation，grasping，deep neural network

要約(英語): We have developed a new loss function focusing on grasping product to automate the display of products by robots. When robots grasp product, it is required to select a grasp product from several products and detect it accurately. In recently years, object detection methods using deep neural network(DNN) has become more common. But even the latest DNN often does not recognize the boundaries of adjacent objects. So, the DNN erroneously detects two products as one product. We address the new problem of correctly detecting adjacent products. To solve the problem, we focus on the fact that the robot grasps only one product at a time, and have developed a new loss function called “mutual loss”. The mutual loss is larger when the DNN erroneously detects adjacent products as one product, so that the DNN correctly detects the products to be selected. Experiments on the OCID dataset showed that IoU of detected object and grasping success rate improved by 6.6pts and 5.8pts, respectively, when DNN uses our mutual loss.

本誌: 電気学会論文誌C（電子・情報・システム部門誌） Vol.143 No.12 （2023） 特集：電気・電子・情報関係学会東海支部連合大会

本誌掲載ページ: 1106-1112 p

原稿種別: 論文／日本語

電子版へのリンク: https://www.jstage.jst.go.jp/article/ieejeiss/143/12/143_1106/_article/-char/ja/