Semantic Based Sport Video Browsing

The current book is a combination of number of great ideas, applications, case studies, and practical systems in the domain of Semantics. The book has been divided into two volumes. The current one is the second volume which highlights the state-of-the-art application areas in the domain of Semantics. This volume has been divided into four sections and ten chapters. The sections include: 1) Software Engineering, 2) Applications: Semantic Cache, E-Health, Sport Video Browsing, and Power Grids, 3) Visualization, and 4) Natural Language Disambiguation. Authors across the World have contributed to debate on state-of-the-art systems, theories, models, applications areas, case studies in the domain of Semantics. Furthermore, authors have proposed new approaches to solve real life problems ranging from e-Health to power grids, video browsing to program semantics, semantic cache systems to natural language disambiguation, and public debate to software engineering.


Introduction
In this chapter, we focus our attention on semantic based sport video highlights detection and semantic based sport video browsing. Users are more interested in sport video highlights than the normal kicks. They site before their TV sets to enjoy the exciting moments that their favorite teams shooting goals. As audiences enjoy the highlight of the video content, usually they have patient to the inserted video Advertisements. For the web based video services, the click rates of sport video highlights are far more than the whole video sequences [56]. Detecting highlights effectively is not only of interest to users/audiences (they are willing to view the highlights) but also of interest to commercial companies. They are interested in inserting their advertisement around the highlights to get more attention from consumers and expand the influences of their products/services. In many online video services websites, the sport video highlights are manually labeled which is very time-consuming and expensive. Thus automatic sport video highlight detection is very urgent [1]- [29]. It is a fundamental step for semantic based video browsing [1,7,8]. However, due to the semantic gaps between computer and human beings, highlights detection is not a trivial.
Two types of approaches have been widely utilized in sport video highlight detection, as shown in Fig.1. The first type of highlights detection approaches are carried out by mapping from low-level features directly or using model-based approaches. In the second type of approaches a mid-level semantic layer is introduced between the low-level features and high-level semantics. Highlights are detected from mid-level semantics rather than from low-level features directly. Thus it is robust against the divergences of low-level features. The second type of sport video highlight detection approaches is more effective than the first type of approaches. Based on the detected high-level semantics, semantics based video browsing can be performed. The conventional video content browsing is based on the summarization of the whole video content. However viewers may be interested to the highlights of sport video. In this chapter, a semantic based sport video browsing approach is introduced. It is a novel book style based approach. Sport video summarization structure is similar to the table of content (ToC) of books. The semantic based video browsing approach has several advantages. Firstly, it is a hierarchical video summarization framework. Secondly, it provides seamless connections with sport video high-level semantics extraction. Thirdly, it is very convenient for users to find their interested content in a large scale database.
In this chapter, firstly learning based soccer video highlight detection approaches are expressed and then semantic based soccer video browsing approach is introduced. The main content of this chapter is organized as follows. Effective low level features representation for sports video is given in Section 2. Middle level semantic classification approaches for soccer video is provided in Section 3. High level semantic detection approaches are introduced in Section 4. Semantic based video browsing is given in Section 5 and conclusions are drawn in Section 6.

Low-level feature representation
In this section, several effective low-level visual features for sport video content analysis are introduced. The features are dominant color, motion, texture and overlaid text information. From the low-level visual feature detection results, mid-level semantics classification and high-level events can be determined by using either direct mapping or statistical learning based approaches.

Dominant color
Dominant color is an effective feature for soccer video analysis. The dominant color is essentially the color of the grass field. The distribution and percentage of grass field region offer significant cues for mid-level semantics categorization and highlights detection.

Related work on dominant color extraction
In [14], Duan et al. classified video shots into eight categories by fusing the global motion pattern, color, texture, shape, and shot length information in a supervised learning framework. Motion and dominant color information of a video shot are fused by multi-layer hidden Markov models (HMM) to determine its categorization [6]. Dominant color ratio and the dominant color projection histogram [12] are utilized for semantic soccer video shot classification. The dominant color extraction approach consists of two stages: dominant color modeling and adaptive dominant color refinement [12]. The dominant color modeling can be viewed as determining a coarse dominant color, which can be utilized for various soccer videos. While the dominant color refinement can be viewed as getting more accurate dominant color for a specific video.
In [56], dominant color detection scheme consists of three steps: 1) Initial dominant color modeling in HSI color space. The accumulative histograms of the HSI components are constructed from training frames of global views randomly selected form wide range of soccer videos and the initial dominant color are determined from the color histogram [36]. 2) Initial dominant color region determination. The initial dominant color and the cylindrical metric [36] are then utilized to classify each pixel of current frame into dominant color or non-dominant color. 3) Adaptive dominant color refinement. For the pixels labeled as initial dominant color, their accumulative histograms of HSI are reconstructed again and the same operations as in initial dominant color modeling step are utilized to obtain the refined dominant color   000 ,, HSI of the current frame. Let's give a brief overview of this approach.

Coarse to fine dominant color extraction approach
The distance of a pixel located at coordinate (i, j) with color   Similar to [36], the dominant color region map DCRM(i, j) as follows: where D th is learned from several soccer video clips, DCRM(i, j)=1 indicates that the pixel (i, j) belongs to the field region. The dominant color ratio (DCR) of a frame is obtained as follows: 11 (, ) where H and W are the height and width of a frame. Fig.2 shows the corresponding dominant color region extraction results. The dominant color region of Fig.2(a) is shown in Fig.2 (b) and the field region is extracted as shown in Fig. 2 (c). The dominant color distribution is represented by the dominant color projection vectors of DCRM.
Based on DCRM, both global or grid based dominant color distributions can be utilized. In [56], the DCRM is parsed into 8 equal-sized regions in both vertical and horizontal direction. By calculating the dominant color pixel ratio of each region, a 16-bin dominant color distribution vector is constructed. In addition to the dominant color distribution, a 255 dimensional block wise color moment generated from 5-by-5 grids of the images is also utilized.
(a) Original image (b) dominant color region (c) field and non-field region Fig. 2. Dominant color region, field region and field lines detection results.

Texture feature
Texture features are insensitive to lighting, rotation, and other conditions. Two kinds of texture features are utilized for coarse shot classification in [56]. The first is hierarchical wavelet packet texture descriptor (HWVP) [47]. The second is 24 dimensional histogram of oriented gradient (HOG) [46]. HWVP descriptor is a 210 dimensional feature which is extracted under local partitioning pattern Local5 (the image is partitioned into 2×2 grids and a centralized grid), by setting the wavelet packet basis to be db2, with hierarchical wavelet packet transform level 2.

Title text detection, localization, and tracking
Title texts provide valuable information for event detection in soccer video content analysis.
Usually the overlaid text about goal, foul, yellow/red card are followed the corresponding highlight events. In Fig.3, overlaid texts (including bulletin texts and title texts) and replays of a soccer video are plotted. There are 20 replays and 7 overlaid texts. The first text is a bulletin text that shows the players' name of the two teams. The second text shows the initial score of two teams. The 3rd and the 6th texts show the names of the players who got goals. The 4th and 7th texts show the updated scores just after a goal. Moreover, the 5th text shows the score of two teams. From Fig.3, the co-occurrences of the title texts and replays always indicate the appearing of highlights. So, it is reasonable to fuse the title text detection results to improve the highlight detection performance. In [56], the detected title text information is utilized for improving highlight detection performances. Thus from the production knowledge overlaid text information is one of the effective clues for high-level semantics inference.
There are four types of texts in the sport video sequences: 1) long term texts, such as icons of the TV channels, the score-boards which appear at the four corners of video frame, as shown in Fig.4(a) and (b) respectively. This type of texts exists in a fixed location in a frame with long duration. 2) Title texts, e.g. showing the score of two teams after getting a goal, as shown in Fig.4(b). 3) Scene texts, such as the texts appearing in signboard and cloth. 4) Bulletin texts, e.g. showing the name list of a team. This type of text usually appears at the beginning of a soccer video. Except utilizing the text detected in sports video, web-casting text information is incorporated with audio-visual features to improve highlights detection performances [28,29].
In soccer video the title texts usually appeared in the bottom-center of the image. The icon, scoreboard and time tables are appeared in the top-left and top-right parts of an image respectively. This type text always appeared during the whole video sequence, for example, the extracted local text lines of Fig. 4(a), as shown in Fig.4(c) are appeared in the whole video. However, the title text as appeared in Fig.4(b), as shown in Fig. 4(d) usually existed in a limited time range. The title text is usually provides more information on its corresponding highlight event inference than a long term text. They are purposively added during soccer video editing, and they are aimed at providing the audience some indicative information about the video content, such as a card or a goal [56], [62].

Related works on video text detection and tracking
The corresponding text detection approaches can be classified into following 3 types. The first is the connection characteristics of video texts [63], [64]. The text detection methods based on this characteristic assume that text regions have uniform colors and satisfy certain constraints on size, shape, and spatial layout. The second is the texture alike characteristic of the text regions [65,33,66,67]. The text detection methods based on texture information usually assume that the text regions have special texture patterns. And the third is the edge density information [34], [68]. These methods make full use of the fact that the edge densities of background are comparatively sparser than those of the text regions [69], [70].
Usually the corner point number in the text region is larger than that in the background regions.
Video text detection and localization can be carried out both in pixel and compressed domains [71,34,68,65,72]. In order to eliminate false detections, the edge [69], texture [33], and shape information [57] are often utilized in text verifications. In addition, the available redundant temporal information is often used in candidate text region verification and falsely detected text region elimination [71,34,68,65,72,73]. Lyu et al. proposed a multiresolution based text detection method [34]. Firstly, original edge map is generated for each of target video frames. Multi-resolution text maps of a target frame are generated by downsampling the original text map. Then text detection, verification and localization are carried out on multi-resolution text maps. Finally, the text detection texts in various resolutions are integrated.

Text detection and tracking
In this chapter, we utilize the corresponding text detection and tracking approaches [56]. The block diagram of the title text detection, localization and tracking is shown in Fig.5. Text line number, spatial location, and temporal duration constraints are utilized during title text detection, localization and tracking. Lyu et al' method [34] is used to detect and localize title texts for a given frame.
It is not necessary to carry out text detection for each of the video frames. In our observations, we find that title texts usually exist for about 5 seconds in soccer video.
Detecting text with an interval of one or half second is enough, e.g. detecting text in intraframes of sport video. Only the texts appeared at the bottom-center of the images and with sufficient duration are determined as candidate title texts. The starting and ending frames of each title text are determined by text line matching and tracking [33].

Motion feature representation
Motion information is also very important for sport video content analysis [14,15,17]. Typically the camera men operate cameras, by fast track, or zoom in to provide audiences with clearer views of the games. Based on the camera motion (i.e. global motion) patterns and domain related knowledge, high level semantics can be inferred.
In [18], global views of soccer video are further refined into the following three types: stationary, zoom and track in terms of camera motion information using a set of empirical rules with respect to domain and production knowledge. The key-frames of a shot with stationary by means of average motion intensity and average motion intensities of global motion.
Global motions in a video sequence are caused by camera motion, which can be modeled by parametric transforms [30,32].
where (,) xy and (,) xy   are the coordinates in the current and the reference image respectively, with the set of parameters denoting the global motion parameters to be estimated.
Average motion intensities of global motion and local motion are utilized for coarse semantic refinement [56]. The average global motion intensity (AGMV ) is calculated as follows: where (,) tt xy  are the warped coordinates in the reference frame by the global motion parameters from the coordinate (x t , y t ).
The local motion information is represented by average motion intensity (AMV ) which is expressed as follows where (MVx j , MVy j ) is the motion vector (MV) of the block with its coordinates (x j , y j ) and j is the block index.

Scene change detection and logo detection
Parsing the sequential video sequences into shots is helpful for video content analysis. This is often called scene change detection or shot boundary detection. Thee sport video sequences are different with other video sequences, e.g. the highlights are often replayed. Usually, logos are utilized to connect the live broadcasted clips with the replayed segments. Thus, scene changes and logos are the basis of sport video semantics detection and semantic based sport video content browsing. Scene change detection and logo detection approach consists of the following three steps [3,56]: (1) logo template detection based on the fact that different starting or ending logo transitions, (2) logo transitions detection in the video program using the logo template based on the probability models of pixel-wise intensitybased mean-square difference and color histogram-based mean-square difference, (3) the replay segments identification.

Mid-level semantic classification
In this section, the corresponding middle level semantics for sport video are defined and detected. Related work on middle level semantic classification is reviewed.

Related work on mid-level semantic classification
Xu et al. classified each soccer video shot into one of the following 3 views : global, zoom-in and close-up [16]. From the view labels, soccer video sequences are further parsed into play and break. Duan et al. classified video shots into predefined mid-level semantics [18]. Based on which, soccer video is coarsely parsed into two type in-play and out-of-play [14]. In [43], global motion and visual features are fused to carry out shot categorization for basketball video. Tan et al. also segmented a basketball sequence into wide angle, close-up, fast break and possible shoot at the basket [61]. Motion and dominant color information of a video shot are fused by multi-layer hidden Markov models (HMM) to determine its category [2]. In [37], each shot of a baseball video is classified into predefined semantic scenes by fusing features extracted from visual content of the image, object level feature representations, and camera motion. In coarse mid-level semantic categorization, soccer video shots are parsed into four coarse views of global, median, close-up, and audience using four one-versus-all SVM classifiers as shown in Fig.7 (a). The kernels of SVM classifiers are RBF (radius-basis-function). The input of the four SVM classifiers is a 505 dimensional low-level feature, including 255d color moment, 16d dominant color distribution, 24d HOG feature, and 210d HWVP feature.

Refinement for coarse mid-level semantics
As median, audience, and close-up views are related to the details, which do not need further refinement. Moreover, the audience shots are rarely replayed, the replay into three types: replay-global, replay-median and replay-close-up using the trained SVM classifiers for coarse semantics classification [56]. The block diagram of global view refinement is shown in Fig.7 (b) respectively.
Global view is further refined into one of the following three types: stationary, zoom and track with respect to camera motion information using a set of empirical rules. A shot is determined as with stationary if AMV<0.5, otherwise non-stationary. The non-stationary shot is further refined into track if m 0 =m 5 =1, otherwise zoom. A zoom is a zoom-in if m 0 =m 5 >1 and a zoom-out if m 0 =m 5 <1. The track is a slow-track if 2 AGMV  , otherwise a fasttrack.
A stationary shot is further refined into stationary-with-goal-post and normal stationary according to the field lines detection results. When the valid field line number is larger than four, then the frame is a stationary with goal-post (SG), otherwise a normal stationary [56].
After mid-level semantic classification and refinement, each segment of a soccer video or event clip is classified into one of 13 mid-level semantics: logo, audience, median, close-up, replay-global, replay-median, replay-close-up, fast-track, slow-track, zoom-in, zoom-out, normal stationary, and stationary-with-goalmouth.

High-level semantic detection
In this section, the high level sport video semantics detection approaches are illustrated in details. Statistical learning models such as hidden Markov model (HMM) as shown in Fig.8(a), enhanced hidden Markov model (EHMM), and hidden conditional random field (HCRF) as shown in Fig.8 (b), based soccer video event detection approaches are described. Correspondingly, soccer video highlight detection results are given.
Xie et al. proposed multi-HMMs based method to improve the performance of play-break detection [6]. The HMMs are different structures and with the same input. Dynamic programming is utilized to fuse the outputs of multi-HMMs to determine the shot type (play or break). In [26], soccer video is segmented into sequential event clips: forward pass, shot on goal, placed kick, turnover, counter-attack, and kick-off using a set of domain rules and sport video production knowledge. The placed kicks are further refined into corner kick, free kick and penalty according to the distribution of players' position.
Wang et al. proposed conditional random fields (CRF) based soccer video event detection method [11]. Firstly, mid-level semantics are determined by mapping from low-level audiovisual features. Then, highlights are detected by fusing the mid-level semantic using CRF [11]. In [27], dynamic Bayesian networks (DBN) are utilized to model goal, corner kick, penalty, and foul events. Low-level audio-visual features are mapped into mid-level nodes of Bayesian networks directly. Bayesian networks model the dependency of the observations of each shot for event type inference. The shot-based event recognition results are fused by DBN to accumulate evidence in the temporal domain [27].
HCRF is utilized to model highlight events in golf and bowling videos [60]. Different from the existing event detection approaches, the transformed low-level features are utilized to perform event detection. Independent component analysis (ICA) is utilized to determine two main components of the input low-level feature. The HCRF is utilized to fuse the ICA components for event type determination.
In most existing works, events are determined from shot level. It is well known that, a single shot separated from its context does poorly in conveying semantics [22], [56]. That is to say the contextual information among the shots of an event clip is not fully disclosed in event inference [11], [27]. Grouping the sequential shots with the same semantics into unified event clips and then recognizing their event types is an optimal way in event detection [56], [21].

Event clip segmentation
Let VS denote a video sequence. Assuming that it consists of K sequential events, the whole video sequence can modeled as follows 1 (,, ) where (1 , , ) t Et K   belongs to one of the predefined events. Each event clip E t is composed of several sequential mid-level semantics, which is expressed as ( We aimed at segmenting VS into event clips by finding event boundaries according to the consistence of semantics. According to domain rules and production knowledge of soccer video, the starting and ending frames of an event clip are the frames at the boundaries where non-global views (including median, close-up, audience, replay and logo) transition to global views [56].

Observations of an event clip
Assuming that an event clip is composed of n mid-level semantics, x j ( 1, , jn   ), the temporal transitions of the mid-level semantics of an event clip can be viewed as observations of statistical learning models. Moreover, in [56], the overall feature extracted from each event clip is combined with the temporal observations for event type inference. The enhanced observation is composed of k overall features ( 3 k  ). They correspond to the title text (TLT), long time break (LTB), and replay (REP) information respectively. All of them are binary. TLT=0 means that there is no title text in event clip. REP =0 denotes that there is no replay in this event clip. LTB=0 represents that the non-global view segment number is very small.

HMM based event detection approach
HMM models the state sequence as being Markov, and each observation being independent of all others given the corresponding state [52]. HMM models the joint distribution under two basic independence assumptions: Markov property and independent property.
In this section the hidden Markov and enhanced Markov models based event detection approaches are illustrated in details. The HMM based approach is carried out event inference using the normal observations of an event clip. While, the EHMM based approach carries out event inference using the enhanced observations of an event clip.

Overview of HMM
HMM is a generative model. It defines a joint probability distribution for the observations and their corresponding labels. HMM models a sequence of observations drawn from a finite state set. HMM carries out inference under two basic independence assumptions [43]. The first assumption is that each state s j depends only on its immediate predecessor s j-1 , and independent of its previous states (i.e. 12 ,, j ss   ). The second assumption is that each x j depends only on the corresponding state s j . Let By assuming that observations are statistical independent, then we have 12 12 (x|s, )

The training of HMM
denote the training data. N is the total number of training event clips. For the k-th training clip, its input is x corresponds to the m-th mid-level semantics in the k-th training clip, and () k y is its event label.
The flowchart of the training of HMMs is as follows. Firstly, the mid-level semantics of each event clip are determined from the extracted low-level features. Then, the Expectationmaximization (EM) is utilized to estimate the model parameters (,,) ii i i AB    . The EM algorithm consists of E-steps and M-steps [43]. The E-step computes the forward and backward probability for the given model, and the M-step re-estimates the model parameters. Given an initial model  , the EM algorithm can find a model t  at t-th iteration www.intechopen.com Semantics in Action -Applications and Scenarios 152 such that its performance is better than the (t-1)-th iteration. That is to say for the K training samples, the EM algorithm makes the following equation hold.
The probability indicates how well the model i  matches the given observations x. The probability of the hidden state sequence S can be expressed as 11

HCRF based event detection
Hidden conditional random fields (HCRF) are discriminative models that generalize the hidden Markov models (HMM) and the conditional random fields (CRF) [52]. HCRF models the state sequence as being conditionally Markov given the observations. Unlike HMM, HCRF is also capable of modeling long range dependencies of the observation.

Overview of HCRF
HCRF uses intermediate hidden variables to model the latent structure of the observations as shown in Fig.8 where y' is a possible label for the observations x.

The training of HCRF
The parameters of HCRF are trained on the training data  is a sequence of inputs, and each y i is the label index of the input sequence x i . The parameters estimations in HCRF is carried out by using the following objective function where e l is the parameter vector of the event e. In this chapter, we have e{Goal, Shoot, Normal Kick, Foul, Placed Kick}.

Highlight detection performances evaluation
In The training steps are as follows: 1) manually label the event clips for the test soccer video sequences. 2) manually label the mid-level semantics. 3) manually label the event boundaries; 4) get the corresponding mid-level semantics for each event clip.
The observations of HMM and HCRF are the temporal transitions of mid-level semantics. The observations of EHMM consist of two parts. The first part is temporal transitions of mid-level semantics which is identical to the observations of HCRF. The second is enhanced observations, including the title texts, long term breaks, and replay information of an event clip [56]. Three overall features are utilized in EHMM. The EHMM and HCRF based event detection approaches are on the basis of event clips and the observation of EHMM and HCRF are mid-level semantics of an event clip.
In order to show the performances of model based event detection approach, in this Section the HMM, EHMM and HCRF based soccer video highlights detection performances are evaluated on seven soccer video sequences. These video sequences are captured from a variety of sources. The total duration of the test video sequences is about 625 minutes. Totally, there are 26 goals, 137 placed kicks, 85 fouls, and 109 shoots. Recall NR, precision NP and F-measure F are used to evaluate objective event detection performances, which are defined as follows: where NC, NM, and NF denote the correctly, missed, and falsely detected event numbers. Moreover, confusion matrix is utilized to show the discrimination of HCRF based event detection approach for the five events.

Semantic based sport video browsing
In this section, a semantic based video browsing framework is proposed. Users can view the video content freely like reading books. The semantic based video content browsing approach is organized as the table-of-content (ToC) of a book. Let's give a brief overlook of the ToC of a book, before illustrating the proposed video browsing method. Fig.11 shows the ToC structure of a book which can be departed into following seven layers: (1) book title (BT); (2) chapter (CH); (3) subchapter (SC); (4) paragraph (PH); (5) page (PG); (6) sentence (ST); and (7) words (WD). With respect to the ToC, readers can know its content very well. They can go straight to the interested content by skipping the irrelevant parts. In order to let readers know the overall content of book, the preface, introduction and summary of a book give the brief illustrations of the book, a chapter and a subchapter.
If sport video content is organized as the ToC of a book, it will be convenient for us to browsing its content. In this chapter, a novel book style based semantic video browsing approach is expressed. TABLE II gives the correspondences of ToC of book and soccer video. In the first layer the book title corresponds to the video category, such as baseball and soccer video. In the second layer, the chapter of a book is similar to a set of events. In the third layer, the sub-chapter corresponds to the detailed information of an event. In the fourth layer, the paragraph of a book is corresponding to the shot of an event. In the fifth layer, the page of a book is similar to the frame of a video. In the sixth layer, the sentence of a book is corresponding to the object in a video sequence. And in the seventh layer, the word of a book corresponds to the pixel. Fig.12 shows the ToC of soccer videos. Soccer video is parsed into following seven events: normal kick, foul, free kick (FK), penalty, corner kick, shoot and goal. The free kick, penalty and corner kick are refined from placed kick using the distributions of players' positions [7]. The fifth layer of the video provides the original video sequences of event. Fig. 11. Table-of-Content of a Book. The target of us is how to carry out ToC based sport video summarization. According to the correspondence of ToC of Book and sport video, a four layer summarization framework is utilized in this chapter for soccer video summarization. In the fourth layer, the abstraction is carried out for the semantic shots by extracting several key-frames. The third layer abstraction (i.e. summary) of a video is the shortened video frames of a specified story unit or event. The summary of the ToC of a video sequences in the third layer is the combination of the summarizations extracted in the fourth layer for the semantic shots. In the sports video, there are also two kinds of semantic shots: global shots and non-global shots. The second layer abstraction (i.e. introduction) of a video is the skimmed video frames of story units or events of a certain class. The first layer abstraction (i.e. preface) of a video is providing the corresponding abstractions of the video sequences. Fig.13 shows an example of the proposed sematic based soccer video browsing approach for a goal event. In the fourth layer, it is the continuous video clips of this event. The third layer shows the extracted 11 key-fames of this event clip, which corresponds to the summary of this event. The second layer shows two representative key-frames, which corresponds to the introduction of this event. The first layer is the most representative frames of this event, namely the preface of this event.

Conclusion
In this chapter, semantic based sport video browsing is introduced. Soccer video high-level semantics detection approaches using hidden Markov models, enhanced Markov models, and hidden conditional fields are expressed in detail and their performances are evaluated. From the detected highlights, a semantic based soccer video browsing approach is proposed. The proposed semantic based soccer video browsing approach carries out video content browsing using a book-like structure.

Acknowledgement
This work is supported in part by National Natural Science Foundations of China (NSFC) No.60903121, No.61173109, and Foundation of Microsoft Research Asia.