Citation Context

.... Those repeated scenes should be summarized with fewer representative scenes. In order to remove the redundant scenes from each set of MSPL, we utilize the latent semantic analysis (LSA) proposed in =-=[26]-=- to measure the similarity between two scenes. The LSA is a widely used technique in information retrieval for analyzing the relationship between words and documents. In our method, a main URC and its...

Citation Context

...enotes a shot and each edge denotes the similarity between the two nodes (shots). Subsequently, to achieve reliable face detection, we perform both the AdaBoost-based [23] and the neural networkbased =-=[24]-=- face detectors to detect faces in each scene. An object is identified as a human face only if both the two face detectors classify it as a face. Moreover, to filter out unreliable faces, if a face do...

Citation Context

... scene have appeared in the preceding scenes, but it is hard to determine the number of the preceding scenes used for role face training. The affinity propagation clustering algorithm was proposed in =-=[20]-=- to cluster the detected faces into relevant groups. However, the approach may still misclassify some different roles into the same cluster, or misclassify some the same role into different clusters. ...

Citation Context

...uously appear for longer than 20 frames, it will be removed from the list of faces. In the face clustering process, we first project each detected face to the local binary pattern (LBP) feature space =-=[25]-=-, then perform the affinity propagation algorithm in [20] for face clustering. Since our method only uses the major roles in a movie to construct the RCN, after the face clustering, the major roles in...

Citation Context

...itive-level methods extract significant video events to represent the original video, which is particularly helpful for domain-specific summarization applications. For example, the method proposed in =-=[4]-=- detects soccer game events, such as goal, referee, and penalty box events, through low-level features, and generates a summary to include these detected events. The method proposed in [5] utilizes co...

Citation Context

...d to measure the importance of the URC based on its SRC set. Since typically, in a movie, the role communities in earlier scenes are usually used to develop the relationships of roles in later scenes =-=[18]-=-, an URC with relatively rich SRC interactions is likely to be an important role community. In our method, an RCN is clustered into different relevant groups, each containing a leading URC and its ass...

Citation Context

...s appearing in each scene. Role identification is an essential step in role-based movie analysis. There exist several research works addressing the role identification problem. The method proposed in =-=[19]-=- uses the detected human faces in the preceding scenes in a movie as training data. After that, each face in the succeeding scenes is associated with one of the roles in the preceding scenes. This app...

Citation Context

...omatically identify important content in a video. According to the types of content used for video analysis, existing video summarization methods can be classified into cognitive-level approaches [2]–=-=[6]-=- and affective-level approaches [7]–[11]. The cognitive-level approaches extract audio-visual features from a video to identify a set of important key frames/shots to represent the whole video. Severa...

Citation Context

... video summarization. I. Introduction V IDEO summarization aims at condensing a full-lengthvideo to a significantly shortened version that still preserves the major content of the original video [1], =-=[2]-=-. Movie summarization is a special class of video summarization. It can be applied for browsing through a movie over a handheld device or a personal multimedia system in a short period Manuscript rece...

Citation Context

...cting the RCN for a movie, we perform scene boundary detection and human-face clustering to generate the role-to-scene co-occurrence matrix. In our implementation, we first use the method proposed in =-=[22]-=- to detect scene boundaries for a movie. Based on [22], the scene detection of a video is transformed to cluster shots into TSAI et al.: SCENE-BASED MOVIE SUMMARIZATION VIA ROLE-COMMUNITY NETWORKS 193...

Citation Context

...atent topic driving models. The physiological responses of humans are exploited in [9] to identify the most entertainment segments of a video to produce a summary. The video player system proposed in =-=[10]-=- adaptively controls the playback speed according to the user behavior. Each user is first asked to fastforward a video and then the player learns the user behavior for adapting playback speed. The me...

Citation Context

...cene is defined to be composed of several consecutive shots taken in the same place. Suppose in a movie the kth scene is denoted by mk. Note, the 180° rule is widely used in video and film production =-=[17]-=-, where multiple cameras are used to capture different actors in the same place. For example, in a talk show, the first camera takes the speaking host at beginning. Then, the producer switches to the ...

Citation Context

...dependent story segments. The method proposed in [13] further extends the concept of role networks [12] by additionally considering the role appearances between adjacent shots. The method proposed in =-=[14]-=- uses an audio segmentation method and the maximum a posteriori probability (MAP) approach to identify the major actors of a movie. The method proposed in [15] uses face clustering and role appearance...

Citation Context

...proposed in [4] detects soccer game events, such as goal, referee, and penalty box events, through low-level features, and generates a summary to include these detected events. The method proposed in =-=[5]-=- utilizes concept-expansion trees to construct a relational graph for characterizing the semantic concepts of documentary videos. A graph modeling-based method is proposed in [6] to detect scene chang...

Citation Context

...d faces into relevant groups. However, the approach may still misclassify some different roles into the same cluster, or misclassify some the same role into different clusters. The method proposed in =-=[21]-=- further combines face clustering and the character name static in the film script to identify roles. As shown in Fig. 3, before constructing the RCN for a movie, we perform scene boundary detection a...

Citation Context

...assigned movies and the corresponding summaries as many times as they want in their own chosen places. In our experiments, we designed the questionnaire based on the set of questions used in [29] and =-=[30]-=-. More discussions about questionnaire-based subjective ranking for video summarization can be found in [31]. Table V lists the seven questions used for subjective evaluation, denoted by Q1–Q7, respec...

Citation Context

...ng the condensed version not comprehensive enough to viewers. Recently, a few movie analysis works adopt both the concepts of social network analysis and role recognition to identify roles in a movie =-=[12]-=-–[15], where the role interactions in the movie are treated as a social behavior, and are then modeled via a network structure. The method proposed in [12] identifies leading roles and role communitie...

Citation Context

...ed for a video, where each node denotes a shot and each edge denotes the similarity between the two nodes (shots). Subsequently, to achieve reliable face detection, we perform both the AdaBoost-based =-=[23]-=- and the neural networkbased [24] face detectors to detect faces in each scene. An object is identified as a human face only if both the two face detectors classify it as a face. Moreover, to filter o...

Citation Context

...tion, video summarization. I. Introduction V IDEO summarization aims at condensing a full-lengthvideo to a significantly shortened version that still preserves the major content of the original video =-=[1]-=-, [2]. Movie summarization is a special class of video summarization. It can be applied for browsing through a movie over a handheld device or a personal multimedia system in a short period Manuscript...

Citation Context

...e condensed version not comprehensive enough to viewers. Recently, a few movie analysis works adopt both the concepts of social network analysis and role recognition to identify roles in a movie [12]–=-=[15]-=-, where the role interactions in the movie are treated as a social behavior, and are then modeled via a network structure. The method proposed in [12] identifies leading roles and role communities by ...

Citation Context

...atch the assigned movies and the corresponding summaries as many times as they want in their own chosen places. In our experiments, we designed the questionnaire based on the set of questions used in =-=[29]-=- and [30]. More discussions about questionnaire-based subjective ranking for video summarization can be found in [31]. Table V lists the seven questions used for subjective evaluation, denoted by Q1–Q...

Citation Context

...n a video. According to the types of content used for video analysis, existing video summarization methods can be classified into cognitive-level approaches [2]–[6] and affective-level approaches [7]–=-=[11]-=-. The cognitive-level approaches extract audio-visual features from a video to identify a set of important key frames/shots to represent the whole video. Several existing cognitive-level methods usual...

Citation Context

...D SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 23, NO. 11, NOVEMBER 2013 Fig. 1. Overview of the proposed movie summarization framework. approach to affect” that is known from the field of psychophysiology. In =-=[8]-=-, an affective scene classification method was proposed to classify affective audio-visual words based on latent topic driving models. The physiological responses of humans are exploited in [9] to ide...

Citation Context

... as to use the user preferences to guide the summary selection. However, before watching a movie, a viewer might not be able to clearly specify her preference for this movie. Hence, a few works [11], =-=[27]-=- designed user interfaces for viewers to interactively specify their preferred content. Most of the existing works extract low-level features, e.g., color, motion, tempo, from the user specified items...

Citation Context

...st applies our scene-level method to remove unimportant or redundant scenes to generate 15-min to 30-min scene-level summary for a movie, followed by a shot selection and/or keyframe selection method =-=[33]-=- to further shorten the summary to the desired length. As a result, with the multilevel approach, a semantically important scene can be represented with much more shots or frames, compared to shot/key...

Citation Context

...nt the whole video. Several existing cognitive-level methods usually use various low-level features, such as color, texture, audio-visual tempo, and motion features, to identify video highlights [2], =-=[3]-=-. Some cognitive-level methods extract significant video events to represent the original video, which is particularly helpful for domain-specific summarization applications. For example, the method p...

Citation Context

...r experiments, we designed the questionnaire based on the set of questions used in [29] and [30]. More discussions about questionnaire-based subjective ranking for video summarization can be found in =-=[31]-=-. Table V lists the seven questions used for subjective evaluation, denoted by Q1–Q7, respectively. In questions Q1–Q6, subjects were asked to rank the movie summaries from best to worst. In Q7, each ...

Citation Context

...gy. In [8], an affective scene classification method was proposed to classify affective audio-visual words based on latent topic driving models. The physiological responses of humans are exploited in =-=[9]-=- to identify the most entertainment segments of a video to produce a summary. The video player system proposed in [10] adaptively controls the playback speed according to the user behavior. Each user ...

Citation Context

... scenes is better than using the relationships between individual roles, as described in the book of The Power of Film, “Drama is a social art form in which a community speaks to itself about itself” =-=[16]-=-. The concept of role-community networks for movie analysis is important since, following a carefully designed scenario, roles in a movie usually form groups/communities to conduct meaningful social b...

Citation Context

...t considering user preference selections and with user preference selections. There is no standard testing procedure for evaluating the performances of video summarization algorithms. As mentioned in =-=[28]-=-, the existing evaluation approaches for video summarization can be classified into intrinsic and extrinsic methods. In extrinsic methods, a video summary is usually evaluated with respect to its impa...

Citation Context

...mparison of the RC Distribution Similarity Performances of the Proposed Method and the Role-Based Method [12] distribution similarity is defined as the histogram intersection of the two distributions =-=[34]-=- as RCsim(pMS, pref ) = ∑N k=1 min(pRC k, prefRC k). (12) In (12), the higher the similarity metric, the more consistent the selected RCs between the automatically generated summary and the reference....