| 講演名 | 2022-09-09 Presentation Slide Assessment System using Visual and Semantic Segmentation Features 易 聖舟(東大), ?上 純一郎(ルバート), 山崎 俊彦(東大), |
|---|---|
| PDFダウンロードページ |  PDFダウンロードページへ |
| 抄録(和) | In this paper, we present a new presentation slide assessment system that can consider structural features of the slides more easily. Our previous work used a neural network to identify novice vs. well-designed presentation slides based on visual and structural features. However, the structural feature extraction was based on the bounding box information of a PPTX file. Therefore, it is unavailable for the users who are unwilling to upload editable PPTX files and those who use other applications such as Google Slides and Keynote. In order to solve this problem, we extract the semantic segmentation of presentation slides from the slide images as a new format of structural features to replace the previous structural features extracted from XML files (i.e., PPTX files). The proposed multi-modal Transformer extracts the visual and structural features from the original images and semantic segmentation results, respectively, to assess the slide design. The prediction targets are the top-10 checkpoints pointed out by the professional consultants. Class-imbalanced learning methods are used for addressing the imbalanced label distribution, and multi-task learning are also applied to improve the accuracy of the proposed model. In the optimal settings of the used machine learning methods for each checkpoint, the proposed model only requiring slide images achieved an average accuracy of 81.67% that is comparative to the performance of the previous work requiring slide images and XML files. |
| 抄録(英) | In this paper, we present a new presentation slide assessment system that can consider structural features of the slides more easily. Our previous work used a neural network to identify novice vs. well-designed presentation slides based on visual and structural features. However, the structural feature extraction was based on the bounding box information of a PPTX file. Therefore, it is unavailable for the users who are unwilling to upload editable PPTX files and those who use other applications such as Google Slides and Keynote. In order to solve this problem, we extract the semantic segmentation of presentation slides from the slide images as a new format of structural features to replace the previous structural features extracted from XML files (i.e., PPTX files). The proposed multi-modal Transformer extracts the visual and structural features from the original images and semantic segmentation results, respectively, to assess the slide design. The prediction targets are the top-10 checkpoints pointed out by the professional consultants. Class-imbalanced learning methods are used for addressing the imbalanced label distribution, and multi-task learning are also applied to improve the accuracy of the proposed model. In the optimal settings of the used machine learning methods for each checkpoint, the proposed model only requiring slide images achieved an average accuracy of 81.67% that is comparative to the performance of the previous work requiring slide images and XML files. |
| キーワード(和) | Presentation Slide / Feature Learning / Class Imbalance / Multi-Task Learning |
| キーワード(英) | Presentation Slide / Feature Learning / Class Imbalance / Multi-Task Learning |
| 資料番号 | MVE2022-13 |
| 発行日 | 2022-09-01 (MVE) |
| 研究会情報 | |
| 研究会 | MVE |
|---|---|
| 開催期間 | 2022/9/8(から2日開催) |
| 開催地（和） | 東京大学 本郷キャンパス＋オンライン開催 |
| 開催地（英） | |
| テーマ（和） | メタバースエクスペリエンスの魅力、メディアエクスペリエンスおよび一般（魅力工学（AC）研究会協賛） |
| テーマ（英） | |
| 委員長氏名（和） | 清川 清(奈良先端大) |
| 委員長氏名（英） | Kiyoshi Kiyokawa(NAIST) |
| 副委員長氏名（和） | 新井田 統(KDDI総合研究所) |
| 副委員長氏名（英） | Sumaru Niida(KDDI Research) |
| 幹事氏名（和） | 磯山 直也(奈良先端大) / 原 豪紀(大日本印刷) / 福嶋 政期(東大) / 後藤 充裕(NTT) |
| 幹事氏名（英） | Naoya Isoyama(NAIST) / Takenori Hara(DNP) / Shogo Fukushima(Univ. of ToKyo) / Mitsuhiro Goto(NTT) |
| 幹事補佐氏名（和） | 宍戸 英彦(筑波大) / 中澤 篤志(京大) / 東條 直也(KDDI総合研究所) / 萩山 直紀(NTT) |
| 幹事補佐氏名（英） | Hidehiko Shishido(Univ. of Tsukuba) / Atsushi Nakazawa(Kyoto Univ.) / Naoya Tojo(KDDI Research) / Naoki Hagiyama(NTT) |
| 講演論文情報詳細 | |
| 申込み研究会 | Technical Committee on Media Experience and Virtual Environment |
|---|---|
| 本文の言語 | ENG |
| タイトル（和） | |
| サブタイトル（和） | |
| タイトル（英） | Presentation Slide Assessment System using Visual and Semantic Segmentation Features |
| サブタイトル（和） | |
| キーワード(1)（和/英） | Presentation Slide / Presentation Slide |
| キーワード(2)（和/英） | Feature Learning / Feature Learning |
| キーワード(3)（和/英） | Class Imbalance / Class Imbalance |
| キーワード(4)（和/英） | Multi-Task Learning / Multi-Task Learning |
| 第 1 著者 氏名（和/英） | 易 聖舟 / Shengzhou Yi |
| 第 1 著者 所属（和/英） | 東京大学(略称：東大) The University of Tokyo(略称：UTokyo) |
| 第 2 著者 氏名（和/英） | ?上 純一郎 / Junichiro Matsugami |
| 第 2 著者 所属（和/英） | 株式会社ルバート(略称：ルバート) Rubato Co., Ltd.(略称：Rubato) |
| 第 3 著者 氏名（和/英） | 山崎 俊彦 / Toshihiko Yamasaki |
| 第 3 著者 所属（和/英） | 東京大学(略称：東大) The University of Tokyo(略称：UTokyo) |
| 発表年月日 | 2022-09-09 |
| 資料番号 | MVE2022-13 |
| 巻番号（vol） | vol.122 |
| 号番号（no） | MVE-175 |
| ページ範囲 | pp.16-21(MVE), |
| ページ数 | 6 |
| 発行日 | 2022-09-01 (MVE) |