Abstract

Although the amount of raw surgical videos, namely videos
captured during surgical interventions, is growing fast, automatic retrieval
and search remains a challenge. This is mainly due to the nature
of the content, i.e. visually non-consistent tissue, diversity of internal organs,
abrupt viewpoint changes and illumination variation. We propose
a framework for retrieving surgical videos and a protocol for evaluating
the results. The method is composed of temporal shot segmentation and
representation based on deep features, and the protocol introduces novel
criteria to the field. The experimental results prove the superiority of
the proposed method and highlight the path towards a more effective
protocol for evaluating surgical videos.