An Instant Approach with Visual Concepts and Query Formulation Based on Users’ Information Needs for Initial Retrieval of Lifelog Moments

研究成果: Chapter in Book/Report/Conference proceedingConference contribution

抄録

Smart devices, such as smartphones and wearable cameras, have become widely used, and lifelogging with such gadgets has been recognized as a common activity. Since this trend produces a large amount of individual lifelog records, it is important to support users’ efficient access of their personal lifelog archives. NTCIR Lifelog task series have studied the retrieval setting as a task called Lifelog Semantic Access sub-task (LSAT). This task is that, given a topic of users’ daily activity or events, e.g. “Find the moments when a user was eating any food at his/her desk at work”, as a query, a system retrieves the relevant images of the moments from users’ lifelogging records of their daily lives. Although, in the NTCIR conferences, interactive systems, which can utilize searchers’ feedback in the retrieval process, have showed the higher performance than systems in automatic manner without users’ feedback in the retrieval process, interactive systems rely on the quality of initial results, which can be seen as results of automatic systems. We envision that automatic retrieval that will be used in interactive systems later. In this paper, therefore, based on a principal that the system should be easy to implement for the later applicability, we propose a method scoring lifelog moments using only the meta information generated by publicly available pretrained detectors with word embeddings. Experimental results show the higher performance of the proposed method than the automatic retrieval systems presented in the NTCIR-14 Lifelog-3 task. We also show the retrieval can be further improved by about 0.3 of MAP with query formulation considering relevant/irrelvant writing about multimodal information in query topics.

本文言語英語
ホスト出版物のタイトルNII Testbeds and Community for Information Access Research 14th International Conference, NTCIR 2019, Revised Selected Papers
編集者Makoto P. Kato, Yiqun Liu, Noriko Kando, Charles L. Clarke
出版社Springer
ページ3-15
ページ数13
ISBN(印刷版)9783030368043
DOI
出版ステータス出版済み - 1 1 2019
イベント14th International Conference on NII Testbeds and Community for Information Access Research, NTCIR 2019 - Tokyo, 日本
継続期間: 6 10 20196 13 2019

出版物シリーズ

名前Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
11966 LNCS
ISSN(印刷版)0302-9743
ISSN(電子版)1611-3349

会議

会議14th International Conference on NII Testbeds and Community for Information Access Research, NTCIR 2019
国/地域日本
CityTokyo
Period6/10/196/13/19

All Science Journal Classification (ASJC) codes

  • 理論的コンピュータサイエンス
  • コンピュータ サイエンス(全般)

フィンガープリント

「An Instant Approach with Visual Concepts and Query Formulation Based on Users’ Information Needs for Initial Retrieval of Lifelog Moments」の研究トピックを掘り下げます。これらがまとまってユニークなフィンガープリントを構成します。

引用スタイル