부산대학교 도서관

Video summarization, with the target to detect valuable segments given untrimmed videos, is a meaningful yet understudied topic. Previous methods primarily consider inter-frame and inter-shot temporal dependencies, which might be insufficient to pinpoint important content due to limited valuable information that can be learned. To address this limitation, we elaborate on a Visual Semantic Self-mining Network (VSS-Net), a novel summarization framework motivated by the widespread success of cross-modality learning tasks. VSS-Net initially adopts a two-stream structure consisting of a Context Representation Graph (CRG) and a Video Semantics Encoder (VSE). They are jointly exploited to establish the groundwork for further boosting the capability of content awareness. Specifically, CRG is constructed using an edge-set strategy tailored to the hierarchical structure of videos, enriching visual features with local and non-local temporal cues from temporal order and visual relationship perspectives. Meanwhile, by learning visual similarity across features, VSE adaptively acquires an instructive video-level semantic representation of the input video from coarse to fine. Subsequently, the two streams converge in a Context-Semantics Interaction Layer (CSIL) to achieve sophisticated information exchange across frame-level temporal cues and video-level semantic representation, guaranteeing informative representations and boosting the sensitivity to important segments. Eventually, importance scores are predicted utilizing a prediction head, followed by key shot selection. We evaluate the proposed framework and demonstrate its effectiveness and superiority against state-of-the-art methods on the widely used benchmarks.