Summarization of human activity videos via low-rank approximation

A salient dictionary learning framework for activity video summarization via key-frame extraction

“…However, the problems of dictionary-of-representatives approaches are especially pronounced when summarizing activity videos, due to their characteristic properties: static camera, static background, heavy inter-frame visual redundancy and lack of editing cuts. This paper, which integrates and extends preliminary work [30] [31], introduces a framework for activity video summarization that attempts to overcome the above issues. Its contributions are four-fold.…”

“…During empirical evaluation of this work, the special characteristics of activity videos were exploited so as to avoid the subjectivity outlined above. Temporal video segmentation ground-truth annotation data, describing obvious temporal boundaries between consecutive activity video segments, were employed for evaluating the proposed methods as objectively as possible, similarly to [30]. Given a summary s of an input video D, the number I s of extracted key-frames derived from actually different activity segments (hereafter called independent keyframes) is used as an indirect indication of summarization success.…”

“…Second, three key-frame extraction algorithms are formulated in this context, where the video reconstruction term is modeled as a Column Subset Selection Problem (CSSP) [2]. This guarantees summary conciseness, favors summary compactness and had not been utilized for key-frame extraction before the preliminary work that this paper extends ( [30]). Third, the saliency terms for the Numerical and for the Greedy algorithms are novel, video-oriented modifications of state-of-the-art image saliency algorithms ( [26] and [10], respectively).…”

High-Level Multiple-UAV Cinematography Tools for Covering Outdoor Events

“…For example, unexpected/unplanned events might occur during live coverage (or even expected events at unknown time-instances), including e.g., leader break-away, crashes, falls, accidents. A number of these events may be automatically detected visually, by on-the-fly activity recognition [37], [38], [39], [40], [41], [42], or activity video summarization [43], [44], [45], [46], [47], [48], [49] systems. Alternatively, they can be manually annotated by a director that overviews the coverage.…”

Deep learning for digitizing, analysing and modelling choreographic 3D signal sequences of intangible cultural heritage

“…In the broad research area of dance summarization, algorithms focusing on extracting key frames of human actions can be also considered. More specifically, the works of [45] and [46] introduce a classification framework for retrieving representative human actions, while the work of [47] proposes a hierarchical union of sub-spaces for human activity abstraction under a semi-supervised framework. In addition, the work of [48] proposes Histograms of Grassmannian Points for classifying multidimensional time-evolving data in dynamic scenes.…”