Video Summarization Based on Feature Fusion and Data Augmentation

During the last few years, several technological advances have led to an increase in creation and consumption of audiovisual multimedia content. Users are overexposed videos via social media or video sharing websites mobile phone applications. For efficient browsing, searching, navigation across collections repositories, e.g., for finding that relevant a particular topic interest, this ever-increasing content should be efficiently described by informative yet concise representations. A common solution problem is construction brief summary video, which could presented user, instead full so she/he then decide whether watch ignore whole video. Such summaries ideally more expressive than other alternatives, such as textual descriptions keywords. In work, summarization approached supervised classification task, relies on feature fusion audio visual data. Specifically, goal work generate dynamic summaries, i.e., compositions parts original include its most essential segments, while preserving temporal sequence. This annotated datasets per-frame basis, wherein being “informative” “noninformative”, with latter excluded from produced summary. The novelties proposed approach are, (a) prior classification, transfer learning strategy use deep features pretrained models employed. These been used input classifiers, making them intuitive robust objectiveness, (b) training dataset was augmented using publicly available datasets. evaluated three user-generated videos, it demonstrated data augmentation able improve accuracy based human annotations. Moreover, domain independent, any extended rely richer representations modalities.