全方位画像のセマンティックセグメンテーションと距離推定の同時学習

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: 2024/06/01

タイトル(英語): Simultaneously Learning Semantic Segmentation and Depth Estimation from Omnidirectional Image

著者名: 横田　敦志（広島市立大学大学院情報科学研究科），李　仕剛（広島市立大学大学院情報科学研究科），神尾　武司（広島市立大学大学院情報科学研究科），小作　敏晴（広島市立大学大学院情報科学研究科）

著者名(英語): Atsushi Yokota (Graduate School of Information Sciences, Hiroshima City University), Shigang Li (Graduate School of Information Sciences, Hiroshima City University), Takeshi Kamio (Graduate School of Information Sciences, Hiroshima City University), Toshiharu Kosaku (Graduate School of Information Sciences, Hiroshima City University)

キーワード: 全方位画像，マルチタスク学習，セマンティックセグメンテーション，距離推定　　omnidirectional image，multi-task learning，semantic segmentation，depth estimation

要約(英語): In multi-task learning, the goal is to improve the generalization performance of the model by exploiting the information shared across tasks. In this paper, we propose a neural network that simultaneously learns depth estimation and semantic segmentation of the environment from omnidirectional images captured by an omnidirectional camera. Our proposed neural network is developed by modifying UniFuse network, which was originally developed for depth estimation from omnidirectional images, to simultaneously learn depth estimation and semantic segmentation of the environment by exploiting the features shared between depth estimation and semantic segmentation tasks. In the experiments, the proposed method was evaluated with the well-known Stanford 2D3D Dataset. High accuracy for the two tasks was not obtained with a single network. However, if either of the two tasks was prioritized in learning, the synergistic effect of the two tasks with shared feature maps would improve accuracy, resulting in better results than a single-task network. It showed the effectiveness of simultaneously learning semantic segmentation and depth estimation from omnidirectional images.

本誌: 電気学会論文誌C（電子・情報・システム部門誌） Vol.144 No.6 （2024）

本誌掲載ページ: 560-567 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/144/6/144_560/_article/-char/ja/