시공간 장면 그래프를 이용한 고밀도 비디오-비디오 검색

초록

Existing text-based video retrieval algorithms have primarily focused on finding relevant videos given short text queries. However, these text queries often struggle to capture the complex semantic structures of real-world long videos, which cannot be readily summarized in a single keyword or sentence. Moreover, the challenge of identifying relevant videos based on a query video has been largely overlooked despite its potential applications in content-based video recommendation and video copy detection. In this paper, we introduce a new video-to-video retrieval problem by defining semantic similarities between videos through dense captions. We then propose a novel algorithm that identifies relevant videos to the query video by matching patterns represented by spatio-temporal scene graphs. Experimental results on the MOMA-LRG dataset, which consists of real-world long videos, demonstrate the effectiveness of the proposed method.

키워드