| 大谷 まゆ | 1561005: D, 中間発表 | 視覚情報メディア | 横矢 直和,加藤 博一,佐藤 智和,中島 悠太 |
|
title: Video Summarization using Textual Descriptions
abstract: Creating a video summary is a labor-intensive process. Video summarization can automate this process by extracting important segments from original videos. However, most prior works do not provide control of the content of a video summary as they create video summaries by sampling video segments based on pre-defined criteria such as representativeness or visual saliency of sampled segments. To generate a video summary that reflects user's intention, we exploit text written by a user and address video summarization by sampling video segments relevant to the content of the text. We developed a video summarization method that extracts video summary based on the similarity between the content of text and videos, and we conducted user-study to evaluate generated summaries. Experimental results revealed the advantage of the text-based method and important factors for video summarization. We also developed a model to predict content-similarity of text and videos to improve our summarization method. language of the presentation: Japanese | |||
| 橋岡 佳輝 | 1551079: M, 2回目発表 | 視覚情報メディア | 横矢 直和,加藤 博一,佐藤 智和,中島 悠太 |
|
title: Two-view simultaneous estimation of camera motion and scene structure using deep learning
abstract: Estimation of camera motion and scene structure from multiple images has been studied. It can be used for augmented reality and autonomous driving, for which robustness is required to use in a variety of scenes. Most methods extract feature points in images and make correspondences between them. However in scenes that only have a few feature points in an image, they may fail in estimation. In this study, especially focusing on the two image case we develop a method to estimate the camera motion and the scene structure in a variety of scenes using deep learning. In this presentation, I introduce current progress and our proposed method. language of the presentation: Japanese 発表題目:2枚の画像を対象とした深層学習を用いたカメラの運動とシーン構造の同時推定 発表概要:近年，拡張現実感技術や車の自動運転などでは，複数枚の画像からカメラ運動とシーン構造の推定を行う手法が使用されており，この目的のため手法の頑健性が求められている．多くの手法では画像上の特徴点を抽出し特徴の対応関係からカメラ運動とシーン構造を推定する．しかし，空や道路など画像上に特徴点の少ないシーンでは，特徴点を検出できず推定に失敗することがある．本研究では特に2枚の画像を対象とし，深層学習の手法を用いて多様なシーンにおいて頑健にカメラ運動とシーン構造の推定を行う手法の開発に取り組む．本発表では今までの取り組みと提案手法について述べる． | |||
| 遠藤 栄典 | 1451022: M, 2回目発表 | 視覚情報メディア | 横矢 直和,池田 和司,川人 光男,山下 宙人 |
|
title: The simulation of spatiotemporal dynamics of brain fluctuation at rest abstract: Researches based on Resting-State Networks (RSNs) measured by non-invasive method explain mainly phenomenological correlation and the mechanistic origin of RSNs is still ambiguous. Thus, precedence researches improved differential equations that represent dynamics between excitatory and inhibitory neurons to include empirical connectivity between regions and attempted to reveal spatial correlation of RSNs measured by functional Magnetic Resonance Imaging (fMRI). However, it is impossible to explain RSNs only by fMRI measuring blood flow variation based on brain activity because information processing in human brain performs in sub second time epochs. For this reason, our research suggests introducing microstates as evaluation criteria of temporal dynamics because of representing basic building blocks of human information processing. Subsequently, simulated neuronal dynamics are evaluated in terms of empirical spatial correlation of fMRI and temporal variation of microstates. As a result, Pearson’s correlation between empirical and simulated spatiotemporal dynamics is moderate. language of the presentation: Japanese 発表題目: 安静時における脳活動の時空間的ダイナミクスのシミュレーション 発表概要: 非侵襲式計測法によるResting-State Networks(RSNs)の研究は相関に基づく現象論的説明が中心でありその生成メカニズムは未だ不明である。そこで、興奮性と抑制性ニューロンのダイナミクスを表現した微分方程式に実験による領野間の結合情報を加えることで、functional Magnetic Resonance Imaging (fMRI)によるRSNsの空間的相関関係を説明しようとする研究が行われてきた。しかし、脳の情報処理は数百ミリ秒単位で行われているため、脳の血流の変化を計測したfMRIだけではRSNsを説明することは不可能である。そこで本研究では、時間的ダイナミクスの評価基準として脳の情報処理の基本単位であるMicrostateを導入し、シミュレーションによる神経活動を実験に基づくfMRIの空間的相関およびMicrostateの時間変化の双方の観点から評価を行った。その結果、時間・空間的ダイナミクスの双方とも中程度の相関関係があることが判明した。 | |||
| 西田 篤史 | 1551073: M, 2回目発表 | 視覚情報メディア | 横矢 直和,萩田 紀博,佐藤 智和,中島 悠太 |
|
title: Detecting Important Persons for Videographers Using Deep Neural Networks
abstract: Important region estimation is an essential technique for applications like video retargeting, which automatically crops a region in a video for small screens, and content-based video compression. Important region estimation methods have been proposed that count detected persons in as important regions, besides the use of visual attention models, since persons are one of the main subjects in the video. However, such methods usually do not distinguish essential persons from passers-by, which results in false positives. In this work, we propose a method to classify a person into an important or non-important person using deep neural networks. In this presentation, I talk about current progress and remaining tasks. language of the presentation: Japanese 発表題目：ディープニューラルネットワークによる撮影者にとっての重要人物の識別 発表概要：小さな画面での視聴のために映像の一部を自動的に拡大する映像のリターゲティングや映像のコンテンツに応じた圧縮などにおいて，映像中の重要領域推定が基盤技術として利用されている．多くの映像において人物は主要な要素のひとつであることから，重要領域推定では視覚的注意モデルなどと合わせて，人物領域を重要領域とする手法が提案されている．しかし，このような手法では多くの場合，すべての人物を重要領域としており，例えば偶然通りかかった人物と映像中の主要な人物を区別することができない．そこで本研究では，ディープニューラルネットワークを用いて映像中の重要人物と非重要人物を識別する手法を提案する．本発表では，現在までの進捗，および今後の展望について述べる． | |||
| 伊藤 俊一郎 | 1551012: M, 2回目発表 | インターネット工学 | 小笠原 司,藤川 和利,小林 和真,門林 雄基 |
|
title: On The Use of Device-Intrinsic Fingerprints against Spear-Phishing E-mail Threats
abstract: Damage from information leakage caused by spear-phishing e-mail is a social problem. S/MIME is an effective, yet insufficient countermeasure. In this work, we focused on canvas fingerprint to substitute for S/MIME's digital signature and considered the utilization as a countermeasure against spear-phishing e-mails. As a result, we showed that senders and recipients who have exchanged e-mails can detect spear-phishing e-mails by using the fingerprint and e-mail's hash value. In order to consider both penetration rate and development support, we implemented a prototype as a Thunderbird add-on. language of the presentation: Japanese | |||
| 湯下 弘祐 | 1551117: M, 2回目発表 | インターネット工学 | 小笠原 司,藤川 和利,小林 和真,門林 雄基 |
|
title:Using Memory Forensics to Detect Infected Devices
abstract: Memory forensics tools make it possible to detect suspicious actions such as code injection on memory, which cannot be detected by antivirus-software. However, analyses using memory forensics tools are not generally used except in cases of criminal investigation because it requires a deep understanding of memory and OS. Therefore, I implemented a tool to detect malicious processes in infected devices by using Volatility Framework and tried to detect infections. In my experiment, I succeeded in detecting 5 malicious processes and 9 malicious actions in 10 infected devices. language of the presentation: Japanese | |||