Citation Context

...feature-based data clustering techniques are unsuitable for video classification because of the semantic gap [19]–[21]. Decision tree classifier is also widely used for supervised data classificatio=-=n [33]��-=-�[35], but it may consist of too many internal nodes which are consequently very difficult to comprehend and interpret. Even after pruning, the decision tree structures induced by the existing machine...

Citation Context

...ication rules for the visual concept nodes at the scene level in the classifier. The posterior probability of a video shot with the feature values being in a video scene can be computed via Bayes law =-=[47]-=-: where and indicate the conditional probabilities for the presence and absence of the video scene , (7) (8)sFAN et al.: ClassView: HIERARCHICAL VIDEO SHOT CLASSIFICATION 77 Fig. 7. Bottom-up procedur...

Citation Context

...nsitive video classification problem: When very large video data set comes into view, efficient video database indexing can no longer be ignored [12]. However, the traditional database indexing trees =-=[13]��-=-�[18], such as R-tree, SR-tree, and SS-tree, are unsuitable for video database indexing and management because of the curse of dimensionality [49]. Video retrieval can be performed in an efficient way...

Citation Context

...ical tree structure of our semantics-sensitive video classifier is derived from the domain-dependent concept hierarchy of video contents and is provided by domain experts or obtained by using WordNet =-=[43]-=-, [44]. Each visual concept node in this classifier defines a specific semantic visual concept which makes sense to human beings, the contextual and logical relationships between the higher level visu...

Citation Context

...r from the following challenging problems. Semantics-sensitive video classification problem: When very large video data set comes into view, efficient video database indexing can no longer be ignored =-=[12]. -=-However, the traditional database indexing trees [13]–[18], such as R-tree, SR-tree, and SS-tree, are unsuitable for video database indexing and management because of the curse of dimensionality [49...

Citation Context

... in video databases [1], [2]. The recent development of content-based video retrieval systems has advanced our capabilities for searching videos via color, layout, texture, motion, and shape features =-=[3]��-=-�[11]. However, these content-based video retrieval systems still suffer from the following challenging problems. Semantics-sensitive video classification problem: When very large video data set comes...

Citation Context

...sual features and high-level semantic visual concepts [22]–[32]. The traditional pure feature-based data clustering techniques are unsuitable for video classification because of the semantic gap [19=-=]–[21]. -=-Decision tree classifier is also widely used for supervised data classification [33]–[35], but it may consist of too many internal nodes which are consequently very difficult to comprehend and inter...

Citation Context

...el visual features and high-level semantic visual concepts [22]–[32]. The traditional pure feature-based data clustering techniques are unsuitable for video classification because of the semantic ga=-=p [19]–[-=-21]. Decision tree classifier is also widely used for supervised data classification [33]–[35], but it may consist of too many internal nodes which are consequently very difficult to comprehend and ...

Citation Context

...e the naive users can exchange their subjective judgments with the database system interactively, the online relevance feedback approach is more suitable for serving a large population of naive users =-=[5-=-], [28], [29]. However, the conventional online rele84 IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 6, NO. 1, FEBRUARY 2004 vance feedback techniques suffer from the following problems when they are applied ...

Citation Context

...ve video classification problem: When very large video data set comes into view, efficient video database indexing can no longer be ignored [12]. However, the traditional database indexing trees [13]�=-=��[18]-=-, such as R-tree, SR-tree, and SS-tree, are unsuitable for video database indexing and management because of the curse of dimensionality [49]. Video retrieval can be performed in an efficient way by c...

Citation Context

...o classify the unlabeled video clips to the known semantic visual concepts defined by the domain-dependent concept hierarchy so that more efficient video database indexing and access can be supported =-=[26]-=-. After the classification, the unlabeled video shots inherit the semantic labels assigned for the visual concept nodes they belong to, thus automatic video annotation is supported by using the widely...

Citation Context

...ode on the classifier and database indexing structure), its feature subspace, feature weights (i.e., importances) and classification rule are predetermined without considering the user’s subjectivit=-=y [50]��-=-�[52]. The same domain-dependent concept hierarchy is used as the inherent database indexing tree structure for video management and holden for all the users. While it is very important to enable real...

Citation Context

...importance for video shot representation and classification for the corresponding visual concept node. Instead of searching the weights from the high-dimensional original feature space (i.e., ) [28], =-=[29], -=-we first use decision tree to obtain the feature subsets, and , for the corresponding visual concept node [33]–[35]. Determining the feature subsets first via decision tree has reduced the search bu...

Citation Context

...12]. However, the traditional database indexing trees [13]–[18], such as R-tree, SR-tree, and SS-tree, are unsuitable for video database indexing and management because of the curse of dimensionalit=-=y [49]-=-. Video retrieval can be performed in an efficient way by classifying the similar videos into the same cluster [10], [11]. Unfortunately, there is a semantic gap between low-level visual features and ...

Citation Context

...xing are normally in high-dimensions [49]. One reasonable solution is first to classify videos into a set of clusters and then to perform the dimension reduction on these clusters independently [41], =-=[42]-=-, the traditional database indexing trees can supposedly be used for indexing these video clusters independently with relatively low-dimensional features. However, the pure feature-based clustering te...

Citation Context

...ase browsing can be supported. III. SEMANTICS-SENSITIVE VIDEO CLASSIFIER Video analysis and feature extraction are necessary steps for supporting hierarchical semantics-sensitive video classification =-=[45]-=-. In our approach, a MPEG video sequence is first partitioned into a set of video shots by using our automatic video shot detection technique. In general, threshold setting plays a critical role in au...

Citation Context

...efficient way by classifying the similar videos into the same cluster [10], [11]. Unfortunately, there is a semantic gap between low-level visual features and high-level semantic visual concepts [22]�=-=��[32]. -=-The traditional pure feature-based data clustering techniques are unsuitable for video classification because of the semantic gap [19]–[21]. Decision tree classifier is also widely used for supervis...

Citation Context

...le for video database indexing and management because of the curse of dimensionality [49]. Video retrieval can be performed in an efficient way by classifying the similar videos into the same cluster =-=[10], -=-[11]. Unfortunately, there is a semantic gap between low-level visual features and high-level semantic visual concepts [22]–[32]. The traditional pure feature-based data clustering techniques are un...

Citation Context

... hierarchical video database browsing because of the lack of efficient video summary presentation structure [12]. In order to support video browsing, some pioneer works have been proposed in the past =-=[36]��-=-�[39]. However, these existing techniques just focus on browsing a video sequence and they did not address how to support the concept-oriented hierarchical video database browsing [40]. A key issue to...

Citation Context

...sed in the past [36]–[39]. However, these existing techniques just focus on browsing a video sequence and they did not address how to support the concept-oriented hierarchical video database browsin=-=g [40]-=-. A key issue to the concept-oriented hierarchical video database browsing is whether the visual summaries found make sense to the naive users and how to interpret the contextual and logical relations...

Citation Context

...ual concepts and enable concept-oriented hierarchical video database browsing [12]. – Query-by-keywords is also used in some content-based video retrieval systems based on manual text annotation [3]=-=, [6]-=-. The keywords, which are used for describing and indexing the videos in the database, are subjectively added by database constructionist without a well-defined structure. Since the keywords used for ...

Citation Context

... high importance for video shot representation and classification for the corresponding visual concept node. Instead of searching the weights from the high-dimensional original feature space (i.e., ) =-=[28], -=-[29], we first use decision tree to obtain the feature subsets, and , for the corresponding visual concept node [33]–[35]. Determining the feature subsets first via decision tree has reduced the sea...

Citation Context

...n the classifier and database indexing structure), its feature subspace, feature weights (i.e., importances) and classification rule are predetermined without considering the user’s subjectivity [50=-=]–[52]-=-. The same domain-dependent concept hierarchy is used as the inherent database indexing tree structure for video management and holden for all the users. While it is very important to enable real-time...

Citation Context

...n an efficient way by classifying the similar videos into the same cluster [10], [11]. Unfortunately, there is a semantic gap between low-level visual features and high-level semantic visual concepts =-=[22]–[-=-32]. The traditional pure feature-based data clustering techniques are unsuitable for video classification because of the semantic gap [19]–[21]. Decision tree classifier is also widely used for sup...

Citation Context

...video databases [1], [2]. The recent development of content-based video retrieval systems has advanced our capabilities for searching videos via color, layout, texture, motion, and shape features [3]�=-=��[11]-=-. However, these content-based video retrieval systems still suffer from the following challenging problems. Semantics-sensitive video classification problem: When very large video data set comes into...

Citation Context

...aracterize video in the database: shot-based and object-based. In this paper, we focus on the shot-based approach because video shots are good choice as the basic unit for video content indexing [36]�=-=��[38]-=-. In order to support more efficient video database management, we classify video shots into a set of hierarchical database management units as shown in Fig. 1. In order to achieve hierarchical video ...

Citation Context

...ree structure of our semantics-sensitive video classifier is derived from the domain-dependent concept hierarchy of video contents and is provided by domain experts or obtained by using WordNet [43], =-=[44]-=-. Each visual concept node in this classifier defines a specific semantic visual concept which makes sense to human beings, the contextual and logical relationships between the higher level visual con...

Citation Context

...s have been done to integrate the online relevance feedback with the inherent database indexing structure, thus the conventional online relevance feedback techniques cannot scale to the database size =-=[53]-=-. The conventional nearest neighbor search is also unsuitable for supporting online relevance feedback because it treats all the visual features with the same importance. If the naive users do not hav...

Citation Context

...re-based data clustering techniques are unsuitable for video classification because of the semantic gap [19]–[21]. Decision tree classifier is also widely used for supervised data classification [33=-=]–[35]-=-, but it may consist of too many internal nodes which are consequently very difficult to comprehend and interpret. Even after pruning, the decision tree structures induced by the existing machine lear...

Citation Context

...d indexing are normally in high-dimensions [49]. One reasonable solution is first to classify videos into a set of clusters and then to perform the dimension reduction on these clusters independently =-=[41]-=-, [42], the traditional database indexing trees can supposedly be used for indexing these video clusters independently with relatively low-dimensional features. However, the pure feature-based cluster...

Citation Context

...in each group is less than a predefined threshold , where is the total number of video shots in the group, and is the dimensions of the discriminating features for the corresponding leaf cluster node =-=[48]-=-. A. Integrated Video Query Our hierarchical video database indexing structure can also support more powerful query-by-example. As mentioned above, the naive users can select two approaches to achieve...

Citation Context

...very important to generate this concept hierarchy atomatically according to the user’s subjectivity, however, it is very hard if not impossible for current machine learning techniques to achieve thi=-=s [54]-=-. Using the common knowledge or the domain knowledge from the experts is a good tradeoff for us to address this hard problem now, obviously, automatic techniques are expected to be provided in the fut...

Citation Context

...eo sequence is first partitioned into a set of video shots by using our automatic video shot detection technique. In general, threshold setting plays a critical role in automatic video shot detection =-=[46]-=-. The thresholds for shot detection should be adapted to the activities of video contents. It is impossible to use a universal threshold that can satisfy various conditions because the thresholds for ...

Citation Context

...109/TMM.2003.819583 1520-9210/02$17.00 © 2004 IEEE nologies need to be developed for indexing, browsing, filtering, searching, and updating the vast amount of information available in video databases=-= [1], -=-[2]. The recent development of content-based video retrieval systems has advanced our capabilities for searching videos via color, layout, texture, motion, and shape features [3]–[11]. However, thes...

Citation Context

...ideo sources: movies, video news, and medical videos. The video scene generation results from two sources are shown in Figs. 8 and 9. For one-level and two-state image classification techniques [26], =-=[30]-=-, their classification accuracy can be achieved higher than 90%. As compared with these traditional one-level andsFAN et al.: ClassView: HIERARCHICAL VIDEO SHOT CLASSIFICATION 79 TABLE I THE AVERAGE P...

Citation Context

...archical video database browsing because of the lack of efficient video summary presentation structure [12]. In order to support video browsing, some pioneer works have been proposed in the past [36]�=-=��[39]-=-. However, these existing techniques just focus on browsing a video sequence and they did not address how to support the concept-oriented hierarchical video database browsing [40]. A key issue to the ...

Citation Context

...MM.2003.819583 1520-9210/02$17.00 © 2004 IEEE nologies need to be developed for indexing, browsing, filtering, searching, and updating the vast amount of information available in video databases [1],=-= [2]. -=-The recent development of content-based video retrieval systems has advanced our capabilities for searching videos via color, layout, texture, motion, and shape features [3]–[11]. However, these con...