Modeling Relations between Objects for Referring Expression Comprehension

• Item type: SIG Technical Reports(1)
• 公開日: 2017-07-12
• タイトル: Modeling Relations between Objects for Referring Expression Comprehension
• タイトル / 言語: en
• タイトル: Modeling Relations between Objects for Referring Expression Comprehension
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: コミュニティQA・言語理解
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_18gh
• 資源タイプ: technical report
• 著者名: Ran, Wensheng; Tian, Ran; Naoaki, Okazaki; Kentaro, Inui
• 著者名(英): Ran, Wensheng; Tian, Ran; Naoaki, Okazaki; Kentaro, Inui
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Referring Expression Comprehension (REC) is the task of pointing out the correct object in an image as corresponding to a given natural language expression. In this work, we improve a previous model of REC by explicitly aligning relations between mentions in the language expression to pairs of objects placed in specific relative positions in the image. Evaluation on the RefGoogle dataset [4] shows that our model outperforms previous work ; we also find that, quite surprisingly, the image features extracted from a pre-trained convolution neural network as used by previous research are not as efficient to REC as automatically recognized category labels.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Referring Expression Comprehension (REC) is the task of pointing out the correct object in an image as corresponding to a given natural language expression. In this work, we improve a previous model of REC by explicitly aligning relations between mentions in the language expression to pairs of objects placed in specific relative positions in the image. Evaluation on the RefGoogle dataset [4] shows that our model outperforms previous work; we also find that, quite surprisingly, the image features extracted from a pre-trained convolution neural network as used by previous research are not as efficient to REC as automatically recognized category labels.
• 書誌レコードID / 収録物識別子タイプ: NCID
• 書誌レコードID / 収録物識別子: AN10115061
• 書誌情報: 研究報告自然言語処理（NL）; 巻 2017-NL-232, 号 2, p. 1-7, 発行日 2017-07-12
• ISSN / 収録物識別子タイプ: ISSN
• ISSN / 収録物識別子: 2188-8779
• Notice: SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc.
• 出版者 / 言語: ja
• 出版者: 情報処理学会