Retrieval is a fundamental challenge within the research community of lifelog and the Lifelog Search Challenge (LSC) has been
an important annual benchmarking activity for interactive lifelog
retrieval systems since 2018. This paper proposes MyEachtra (/maiAK-truh/), a system designed for the upcoming LSC’23 workshop.
Improved upon MyScéal, which was the top performing system
from LSC’20 to LSC’22, MyEachtra includes modifications to address the challenges of non-owner user understanding of lifelog
contexts and open-ended lifelog question answering. Specifically,
MyEachtra shifts the focus from images to events as retrieval units.
Events are segmented using location metadata as well as visual
and time differences between successive images. A pilot study
on different approaches to aggregate images into events was conducted to test the automatic performance of the system, which
showed promising results. For known-item queries, showing only
the top 3 events proved to be adequate to find relevant images.
However, future evaluation of the performance for ad-hoc and
question-answering queries is necessary for a complete analysis of
the MyEachtra.
Metadata
Item Type:
Conference or Workshop Item (Paper)
Event Type:
Workshop
Refereed:
Yes
Uncontrolled Keywords:
lifelog, interactive retrieval system, pretrained models, user modeling