JP2002506255A - Method and system for generating semantic visual templates for image and video verification - Google Patents

Abstract

(57) [Summary] For a database image / video search, a semantic visual template (SVT) is a set of icons of exemplary scenes / objects that characterize a concept, eg, skiing, sunset, etc. SVT provides two-way interaction between user and system. The user provides the system with the initial sketch or example image as a seed for automatically generating another representation of the same concept. The user can then select a view of the inclusion that is likely to be reliable to represent the concept. Once the SVT has been established, the database can be searched for the user to provide appropriate feedback on the returned results. Once an SVT has been established, the established SVT allows a user to interact with the system at a concept level. When forming a new concept, an existing SVT can be used. Restricted vocabulary can be parsed by semantic visual templates that query the system.

Description

DETAILED DESCRIPTION OF THE INVENTION

The present invention relates to a technique for searching for still images, videos, and audios in a database, and more particularly to a technique for facilitating access to items in the database.

[Industrial applications]

[0002] Tools and systems for exploring and retrieving visual information have become important as image audio video is formed, disseminated, and stored in digital form.
However, while the effective "search engine (search or search engine)" for text data is widely available, related tools for searching visual image and video data are still underdeveloped.

[0003]

[Prior art]

Typically, images are indexed and searched using keyword technology to obtain image and video databases. Known search systems of this kind have a number of drawbacks, since images are not relevant to text-shaped information, manual inclusion of captions is time-consuming and nominative, and text-shape announcements are usually accompanied. Fails to represent the visual essential features of the scene. For example, a textual description such as "A man is leaning against a brick wall" conveys some visual information about either the man or the brick wall. Such visual information is often important in searching for a particular video.

In recent years, researchers have begun to explore new forms of image and video storage retrieval.
Such a search is based on visual attributes such as color, texture, shape, and similarity of spatial and temporal relationships among the objects making up the video. In this representative example, a query is identified by providing an exemplary image and visual sketch. Thus, the search system retrieves and selects the images and videos with the highest similarity to a given example or sketch.

[0005]

[Problems to be solved by the invention]

Index the database with a collection of visual templates to easily retrieve images and videos from the database. Preferably, according to an aspect of the present invention, the visual template represents a semantic concept or category, eg, skiing, sunset, etc. This forms an architecture that uses semantic visual templates (SVT), each of which describes the concept well.

[0006] A semantic visual template (hereinafter also referred to as SVT) is established by interactive processing between a user and a system. The user can provide a system with an initial sketch or example image as a seed to a system that will automatically generate another representation of the same concept. Thus, the user can pick up those scenes, including the most reliable ones that represent the concept. When the SVT is established, the database is searched by the SVT, and the user is provided with a search capability (relevance) fed back to the selected result.
The established SVT allows users to interact at the concept level as a system. To create a new concept, an existing SVT can be used.

Further, according to the present invention, there is provided a technique for parsing a limited vocabulary of words in relation to a semantic visual template querying system.

[0008]

[Means for Solving the Problems]

The computerized method for generating a visual template for a concept according to the present invention is a computerized method for generating a visual template for a concept, comprising the steps of: a. Obtaining at least one initial query for the concept; b. Generating an additional query; c. Generating an additional test query for relevance for said concept; d. Including an additional template in the visual concept template if relevant. It is characterized by doing so.

[0009]

【Example】

The invention will be described with reference to the drawings. Techniques for computer-based spatial and temporal visual search using visual templates to search images and videos in different categories are disclosed in US Provisional Patent Application No. 60 / 045,637 (May 5, 1997) and PCT International Application No. PCT. / US98 / 0912
4 (May 5, 1998) (Canada, Japan, Korea and the United States). Video Q
Such search techniques, referred to as, can be used in connection with the present invention.

In a video stream, the beginning and end of a scene can be determined using video Q, and can be used, for example, to correct camera motion to extract scene objects. Even with video Q, each object can be characterized by silent attributes such as, for example, body, texture, size, shape and motion. This makes it possible to obtain a video object database consisting of all the objects extracted from the scene and its attributes.

Visual Template A visual template represents the idea of the shape of a sketch or an animation sketch. Since a single visual template is hard to represent the class of interest,
A library of visual templates containing expression templates of different semantic classes can be incorporated. For example, when searching for a video clip of the class "Sunset", one or more visual templates associated with the class are selected and a similarity-based query is used to find the visual template of "Sunset".

An important advantage of using a visual template library is that it links a low-level visual feature display to a high-level semantic concept. For example, if the user enters a forced natural language shape query, as described in the prior patent application above, the natural language query was identified by visual attributes and constraints using a visual template. Can be converted to automatic queries. When indexing the visual content of a repository or database as text, the search method cannot be directly applied to ordinary text.

Semantic Visual Template (SVT) A semantic visual template is a set of visual templates associated with a particular semantic. This SVT announcement has certain key characteristics:

[0014] Semantic visual templates are common in nature. Consider a set of visual templates that better cover a concept for a given concept. Examples of successful SVTs are sunset, high jump and downhill skiing.

Although the semantic visual template for a concept is small,
High Precision-Covers a large percentage of the collection's associated images and videos for recall performance.

The procedure for finding semantic visual templates of different concepts is systematic, efficient and robust. Efficiency is about focusing on a small set of visual templates. Robustness is manifested by applying a new library of templates to new image and video collections.

In connection with Video Q, a semantic visual template can be further understood as a set of icons or exemplary scenes / objects representing the semantics to which the template relates. Extract the feature vector for the query from the semantic visual template. The icon is an anime sketch. In video Q, the features associated with each object and their spatial and temporal relationships are important. Histograms, textures, and structural information are examples of global features that can be part of such templates. Icon-based recognition and the choice between feature vector pairs formed from all global features depends on the semantics to be represented. For example, a sunset scene is best represented by a pair of objects, while a waterfall or crowd is better represented using a global feature set. Thus, each template contains multiple icons, i.e., exemplary scenes / objects. Accordingly, the elements of this set have overlapping application areas. Hopefully this coverage will be minimized with a minimal set of templates.

Each icon for a concept, for example, downhill skiing, sunset, beach crowds, is a visual representation of graphic objects similar to the actual objects in the scene. Each object is associated with a set of visual attributes, eg, color, shape, texture, motion. Each attribute to the concept and the suitability of each object is also specified. For example, for "sunset", the colors and specific structures of objects such as the sun and sky are more relevant. For sunset scenes, the sun object is optional. The reason is that it is a sunset video where the sun is not visible. For the concept "high jump", the motion attributes of the foreground object are mandatory, but the texture attributes of the background are not mandatory, and both attributes are more appropriate than the other attributes. Certain concepts require only one object and represent global attributes of the scene.

FIG. 5 shows several possible icons for “high jump” and FIG. 6 shows several possible icons for “sunset”. The optimal set of icons needs to be selected based on appropriate feedback and maximum coverage from a recall point of view, as described in more detail below.

An efficient technique for generating semantic visual templates for various concepts has been invented. Each semantic concept has several display visual templates, which are used to retrieve significant portions of images and videos for aggressive coverage or high recall from the repository. Aggressive coverage sets of different visual templates can overlap. It is therefore an object to find a small set of visual templates with aggressive coverage that largely overlaps.

A user can provide initial conditions for a valid visual template. For example, a user can use a yellow circle (foreground) and a bright red rectangle (background) as an initial template to search for a sunset scene. The user can also answer the interactive questions to indicate the weight and suitability of different objects and the requirements associated with attributes and context.
This question is responsive to the current query that the user sketched on, for example, the sketchpad.

By providing the match of the initial visual template and all visual attributes of the template, the search system returns a set of most similar images / videos to the user. By providing the returned result, the user can obtain a subjective evaluation of the returned result. The accuracy of the results and the aggressive coverage, ie recall, can be calculated.

The system determines an optimal strategy for modifying the initial visual query and generates a modified query based on the following: 1. 1. suitability of each visual attribute obtained by user's question; 2. the precise recall performance of the previous query, and Information related to the image and video feature level distribution of the repository.

Such a feature is implemented in the technique conceptually illustrated by FIG. 1 with a specific indication of the query for the concept “high jump”. The query includes three objects: two static rectangular reminders and one object that moves for low light. For each object in the query,
The four quantities are specified in relation to weight, for example, color, texture, shape and size, as shown by the vertical bars in FIG. A new query can be formed by stepping on at least one of these quantities, at which point a template interaction can be sought for the template to determine the likely reliable inclusion. Once a suitable number of icons have been incorporated into the temporary template, the template can be used for database searches. The results of this search can be evaluated against recalls and schemes. If acceptable, the template can be stored as a semantic visual template for "high jump".

Template Metric The basic video data comprises a number of segmented objects and is called a video shot. The lifetime of any particular video object is equal to or shorter than the duration of the video shot. A similar measure D between multiple SVT sets and video shots is defined by: D = min {ω _f・ Σ _{i} d _f (O _i , O ′ _i ) + ω _s・ d _s } (1) where O _i is an object defined by the template, and O ′ _i is O _{i is the} aligned object for _i , d _f is the feature distance between its arguments, and d _s is the similarity between the spatio-temporal structure of the template and the spatio-temporal structure between the aligned objects of the video shot. And ω _f and ω _s are scaling weights for feature distance and structure dissimilarity. The query procedure needs to generate a candidate list for each object in the query. Thus, the distance D is minimal over all possible sets of registered objects that satisfy the space-time constraint. For example, if the semantic template has three objects and two candidate objects are retained for each single object query, at most 8 potential objects considering the minimum distance calculation of Equation 1 There is a set of candidates.

Given a query with N objects, a search must be performed over the entire set of N objects that appear together in the video shot. However, the following more economical procedure can be adopted for the economy by using a computer. 1. Query the database of all objects using each video object, i.e., O _i, and use the threshold to obtain a list of matching objects that can be kept short. Then only the objects included in this list can be considered as candidate objects that match O _i . 2. The candidate objects in this list are then combined to form the final set of matching objects that confirm the spatial-temporal structure relationship.

A two-way interaction is used between the template generation user and the template generation system. Given the initial scenario, use the appropriate feedback, and use the technical scope of a small set of icons that gives the greatest recall with such techniques. The user provides the initial query as a sketch of the concept to generate a template consisting of objects with spatial-temporal restrictions. The user can also specify whether the object is mandatory. Each object has the characteristic that the user assigns appropriate weights.

The initial query can be viewed as a point in a higher-order feature space to which all videos in the database can be mapped. In order to automatically generate a set of test icons, it is necessary to jump to each of the features of each of the objects after quantizing the space. To perform the quantization, the size of the space can be determined using a weight specified by the user in conjunction with the initial query, which weight is a measure of the degree of suitability of the attribute attributed by the user to the features of the object. It can be regarded as a guide. Therefore, when the following equation is satisfied, the quantization becomes coarse when the weight is low, or vice versa. Δ (ω) = I / (a ・ ω + b) (2) where Δ is the jump distance corresponding to the feature, ω is the weight associated with the feature, and a and b
Is a parameter, which is chosen such that △ (0) = 1 and △ (1) = d ₀ , which is a system parameter for the threshold set at 0.2 in a typical system . By using the jump distance, the feature pace is quantized into a hyper rectangle. For example, in conjunction with カ ラ ー (ω) for color
Using the metric for the LUV space, one can generate a cuboid.

In order to prevent the total number of possible icons from increasing rapidly, the change of the combination of features is for example prevented as follows. 1. For each feature of the object, the user selects a possibly reliable set of this feature. 2. Thus, the system makes a connection to the set of features associated with the object. 3. Thus, the user selects the bond that most recently represents the change in the object.

In the case of multiple objects, an additional join to the candidate list for each object can be included in step 2. Once a list of potentially reliable scenes is generated, the system is queried using the icon selected by the user. If the user labels the returned result as positive or negative with appropriate feedback to the returned result, determine the icon that was the largest recall.

Concept Scopes For a user to search a database, many scopes are required for a concept.
Each range can belong to a different feature space, for example, describing "sunset" at the object level as well as at the global level. The object-level description may be a set of two objects, such as sky and sun. These objects may be further quantified using signature level descriptions.

As shown in FIG. 2, one concept (eg, sunset) has two different types of states, required (N) and sufficient (S). A semantic visual template is a sufficient condition in the concept, not a requirement, and it is not necessary for a particular SVT to include the concept to its fullest extent. Additional templates may be generated manually, ie, as user-entered additional queries. This task is performed for each concept. The requirements can be given to the concept so that additional templates can be automatically generated once the initial query template is given.

The user interacts with the system via a “conceptual questionnaire” to specify the requirements for the meanings to be searched. These conditions may be global, for example global color distribution, relative spatial and temporal correlations, and the like. Once the necessary and sufficient conditions for the concept have been established, the system moves through the feature space and generates additional templates with the user's original template as a starting point.
This occurrence is also modified by relevant feedback provided by the user to the system. By analyzing the relevant feedback, new rules belonging to the requirement can be determined. These can be used to further modify the template generation procedure. The rules are generated by finding correlations between conditions that are deemed necessary for a concept by videos marked as appropriate by the user. Rules (or relationships)
This principle of determining the "data mining" described in the above-mentioned patent specification,
S. Brin et al., "Dynamic Item Set Counting and Relationship Rules for Shopping Basket Data" ACM SIGMOD Conference on Data Management, 1997, 2.
55 to 246 pages, A paper by Brin et al., "Beyond the Shopping Basket: Generalization-Related Rules for Modification," ACM SIGMOD Conference on Managing Data,
1997, pp. 265-276.

Example of Rule Occurrence A query regarding "crowd" in the video Q is in the form of a sketch. The user specifies a visual query for the object and gives weights regarding color and size, but specifies a composition (of said crowd) or a more detailed description in the form of relative spatial and temporal movement that characterizes the crowd concept It is not possible. However, because the user feels that the concept of "crowd" is strongly characterized by composition and the relative spatial and temporal arrangement of humans,
These are listed as requirements.

Through the process of feedback, the system specifies video clips related to concepts of interest to the user. Here, the system knows that the configuration and spatial and temporal arrangement are necessary for the concept,
Attempts to determine a matching pattern among the features deemed necessary in the relevant video. The patterns are then returned to the user asking if they match the concept they are looking for. If the user accepts these patterns as consistent with the concept, they are used to generate a new query template as shown in FIG. Including this new rule has a dual effect on query template generation, i.e., improves search speed and improves the accuracy of returned results.

Generating a Concept Range The query defines a feature space in which to search. The feature space is defined by attributes of the visual template and associated weights. In particular, the attributes define an axis of the feature space, and the associated weights stretch / compress the associated axis. Within this complex feature space, each video shot can be represented as a point in this space. The visual template includes part of this space. Since the visual templates may differ in features and characteristics (global vs. object level), the spaces defined by the templates are different and do not overlap.

Although the selection of several features may be insufficient to determine a concept,
For example, different suitable choices for weights may be sufficient.
In this way, a certain concept can be mapped to a certain feature space.

A concept is not limited to a certain feature space or a certain cluster. For example, with respect to the sunset video train class, sunset cannot be completely characterized by one color or one shape. Thus, it is important to determine not only the overall static features and weights for a concept, but also those features and weights that may change.

A search for a concept starts by specifying a particular global constant. The contextual questionnaire determines the number of objects in the search and the overall characteristics required for each object. These represent constraints in the search process that do not change.

The user provides an initial query that specifies features and sets weights. The set intersection is obtained by a set of requirements defined by the user. Said requirements remain unchanged. The template is varied based on changes to those features that are deemed sufficient. If the sets are not common, a rule characterizing the concept is obtained based on the requirements and associated feedback.

The associated weight of each feature indicates the tolerance that the user wants each feature to follow. This tolerance is defined as the distance threshold along which each feature follows, eg, d (ω)
= 1 / (a · ω + c), and defines a hyper-ellipse in the searched feature space. The threshold determines the number of possible ranges that do not overlap. The number of ranges determines the size and number of jumps along this particular feature. The algorithm performs a breadth first search and is governed by three criteria.

First, the greedy algorithm goes in the direction of increasing recalls, calculates all possible initial jumps, converts each jump into a corresponding visual template, executes the query, Match the results, show the results to the user for relevant feedback, and select those results that maximize incremental recall as possible points for subsequent queries.

Second, the log search is searched as a subsequent query by obtaining smaller jumps in local regions. The rationale is that the current query point produces good results and should be carefully searched for other templates. The search stops when enough visual templates have been generated to cover as much as 70% of the concept (i.e., 70% or more recalls).

Third, the breadth-first search often uses a feature level distribution to guide the search, as it often creates too many possibilities to perform at once. The distribution along each feature is calculated in advance. Using this information, a jump to a region with a high concentration of video shots is selected.

Language Integration with SVT Text-based queries for known images and videos are based on matching keywords with images or videos. Keywords with data can be generated manually or obtained by association, i.e., taken from attached text (for images) or from titles with video.

Such an attempt excludes the possibility of a practical system containing a very large database of videos or images for the following reasons. It is not possible to manually generate annotations to an existing video database. Many videos do not include a title. There is no instantaneous correlation between the attached title and the video. For example, during a baseball game, the commentator may talk about Babe Ruth's achievements that do not exist during the game being played, and keywords based on text about videos that include "Babe Ruth" may incorrectly display this video. become.

Generating linguistic content for video by analyzing only the video stream reaches a computer vision problem that has been perceived as difficult. A more practical approach is to use the descriptive power of natural language and simultaneously use visual content such as object movement, attributes such as color and texture.

The user types into a string and the system breaks down into a video model. Video Q gives the "language" to enter the query by sketch. As shown in Table 1, there is a simple correlation between what is in video Q and its natural language objects. Table 1 Attribute NL type mobile verb color, texture adjective shape noun spatial / temporal prepositions / conjunctions

A conditional language set can be used with an acceptable set of words. Sentences are parsed into classes such as nouns, verbs, adjectives and adverbs to generate a moving model of the video sequence.

For example, when analyzing the sentence “The building walked slowly toward the sunset,” the system can analyze as shown in Table 2. Table 2 words NL type walking verb slowly adverb towards preposition sunset noun

For verbs, adverbs, adjectives and prepositions, if they are modifiers (or descriptive words) to nouns (objects), a small but fixed database can be used. A database of nouns (ie, scenarios / objects)
Initially, it contains about 100 scenes and can be expanded by dialogue with the user.

Each object can have a shape description that is modified by various modifiers, such as adjectives (color, texture), verbs (walked), adverbs (slowly). It is inserted into the Video Q Palette, where it can undergo fine processing.

If the decoder encounters a word that does not exist in its modifier database (ie, the database corresponds to a verb, adverb, preposition, or adjective, respectively), the decoder looks up synonyms and uses the scenario for that word. To determine if are present in the database and use them instead. If not, the decryptor returns a message indicating an invalid string.

If the decryptor encounters a word that cannot be classified, the user needs to modify the text, or if the word is a noun (such as Bill),
In this case, the noun) is communicated to the system, and additionally that the word is about a human. If the user indicates a noun that does not exist in the system's database, the user attempts to retrieve the sketchpad object so that the system can learn about the object. In the database,
Attributes such as movement, color, texture, and shape can be generated at the object level, so that the alignment level can be that level.

As shown in FIG. 4, another source of information can be an audio stream with video. If audio correlates closely to video, audio can be the single most important source of video semantic content. From the audio stream, a set of keywords can be generated, for example, 10-20 keywords per video sequence. In this case, the search at the keyword level can be combined with the search at the model level. These videos are ranked in the highest rank that matches at the keyword level (for the language) and the mobility model level.

[0056] language visual template 1 to play the video image of the implementation example slalom skiing ya. The system asks questions about the context and the user answers. Label the linguistic visual template "Slalom". This query is specified as an object reference containing two objects. 2. The user sketches the initial query. The large blank background represents the slope of the ski, and the small foreground object represents the user's skier with a unique zigzag trajectory. 3. The highest relevance weight is assigned to all features associated with the background off-ject skier. Background features are identified as static, and skier features may change during template generation. 4. The system generates a set of test icons that allow the user to select the skier's color and change of trajectory characteristics. 5. Combine the four selected colors and the three selected trajectories to form twelve possible skiers. Combine the list of skiers into a single background and get the twelve icons of FIG. Here, it is understood that three adjacent icons have the same color. 6. The user selects a candidate set and queries the system. The system plays the 20 closest video images. The user feeds back the relevance and leads the system to a small set of slalom skiers.

Evening . A database containing 72 sunsets in 1952 video images was used. Using the initial sketch without a verbal visual template, 10% reproducibility and 35% accuracy were achieved. Using the linguistic visual template, 8 icons were generated and 36 sunsets were found with 50% reproducibility and 24% accuracy using these icons.

High jumper . The database contains 9 high jumpers with 2,589 video images. Without a verbal visual template, the recall was 44% and the accuracy was 20%. 56% recall with linguistic visual templates
And the accuracy was improved to 25%. A system focused on a single icon different from the original sketch is provided by the user.

[Brief description of the drawings]

FIG. 1 is an illustration of an interactive technique for generating a library or collection of semantic visual templates according to a preferred embodiment of the present invention.

FIG. 2 is an explanatory diagram showing a concept having necessary and sufficient conditions.

FIG. 3 is an explanatory diagram illustrating query generation.

FIG. 4 is an explanatory diagram of an interactive system according to another preferred embodiment of the present invention including audio processing.

FIG. 5 is an illustration of an initial sketch of a set of icons illustrating the concept “high jump”.

FIG. 6 is a diagram illustrating an initial sketch of a set of icons illustrating the concept “sunset”.

FIG. 7 is an explanatory diagram showing a semantic visual template of the concept “slalom”.

────────────────────────────────────────────────── ─── Continuing on the front page (72) Inventor Sea-Fu Chan 10027 New York Riverside Drive 560 Apartment 18 New York 18 K 34 Ai (72) Inventor Hari Sandaram 10027 New York, New York, United States West One Handed Twenties Street 434 Apartment 9D F Term (Reference) 5B075 ND06 NK02 NK46 PP02 PP13 PQ02 QM08 5C052 AB04 AC08 DD04 EE03

Claims (24)

[Claims] 1. A computerized method for generating a visual template for a concept, comprising: a. Obtaining at least one initial query for the concept; b. Generating at least one additional query for this initial query. C. Generating an additional query for testing for relevance for the concept; d. Including, if appropriate, an additional template in the visual template for the concept. A computerized method for generating visual templates for concepts. 2. The computerized method for generating a visual template for a concept according to claim 1, wherein each query is an icon / example image. 3. The method according to claim 1, wherein the initial query is obtained via a sketchpad. 4. The computerization of claim 1, wherein the additional query steps a query feature having a step size inversely related to a weight associated with the query feature. Method. 5. The computerized method for generating a visual template for a concept according to claim 1, wherein generating an additional icon comprises forming a joint of likely feature values. 6. The computerized method for generating a visual template for a concept according to claim 1, wherein the adequacy is confirmed by interactive user interaction. 7. A computerized method for querying a concept video database using a subset of natural language associated with a semantic visual template, comprising: a. Obtaining a texture query; b. Parsing the query to obtain visual attributes. Using a visual attribute to form a visual query; d. Retrieving information using the visual query; and e. Displaying information. Computerized method of querying a concept video database 8. The computerized method for querying a concept video database according to claim 1, wherein said texture query is obtained from a keyboard. 9. The method of claim 7, wherein the natural language subset comprises a small set of nouns, verbs, prepositions, adjectives, and adverbs. 10. The method of claim 7, further comprising the step of interactively expanding the subset. 11. The parsing step includes: (i) establishing a match between the query and the natural language subset; and (ii) labeling different parts of the query as nouns, verbs, adjectives or prepositions. And (iii) seeking an explanation of whether the words of the query are not present in the natural language subset and labeling the words accordingly. A computerized way to query video databases. 12. The step of forming a visual query comprises establishing a match between a natural language subset and a set of semantic visual templates generated by the method of claim 1, wherein the semantic visual template comprises: Acts to establish the nouns of the query, adjectives that act to modify the visual examples of these nouns, verbs that act to perform the actions, and the spatial and temporal order needed to form the visual query. The computerized method for querying a concept video database according to claim 7, wherein the visual example is a visual example of a preposition. 13. A computing system for generating a visual template for a concept, comprising: a. Means for obtaining at least one initial query for the concept; b. Means for generating at least one additional query for this initial query. C. Means for generating additional test queries for relevance for said concept, and d. Means for including additional templates in the concept visual template if appropriate. A computerized system that generates visual templates for concepts characterized by: 14. The computerized system for generating a visual template for a concept according to claim 13, wherein each query is represented by an icon / example image. 15. The computerized system for generating a concept visual template according to claim 13, wherein the means for obtaining the initial query comprises a sketchpad. 16. The concept of claim 13, wherein said means for generating an additional query comprises means for stepping a query feature having a step size inversely related to a weight associated with the query feature. Computerized system that generates visual templates for 17. The computerized system for generating a visual template for a concept according to claim 13, wherein the means for generating additional icons comprises means for forming a joint of likely feature values. 18. The computerized method for generating a visual template for a concept according to claim 13, wherein the adequacy is confirmed by interactive interaction between a system and a user. 19. A computerized system for querying a concept video database using a subset of natural language associated with a semantic visual template, comprising: a. Means for obtaining a texture query; b. Parsing the query to obtain visual attributes. Means for generating information; c. Means for using visual attributes to form a visual query; d. Means for retrieving information using the visual query; and e. Means for displaying information. A computerized system that queries a video database for a concept. 20. The computerized system of claim 19, wherein the means for obtaining the texture query comprises a keyboard. 21. The natural language subset comprises a small set of nouns, verbs, prepositions,
20. The computerized system for querying a concept video database according to claim 19, comprising adjectives and adverbs. 22. The computerized system of claim 19, further comprising means for interactively extending the subset. 23. The parsing means includes: (i) means for establishing a match between the query and the natural language subset; and (ii) means for labeling different parts of the query as nouns, verbs, adjectives or prepositions. 20. The concept of claim 19, further comprising: (iii) means for asking whether the words of the query are not present in the natural language subset and labeling the words accordingly. A computerized system that queries video databases. 24. The means for forming a visual query comprises means for establishing a match between a natural language subset and a set of semantic visual templates generated by the method of claim 13, wherein the semantic visual template Acts to establish the nouns of the query, adjectives that act to modify the visual examples of these nouns, verbs that act to perform the actions, and the spatial and temporal order needed to form the visual query. 20. The computerized system for querying a concept video database as claimed in claim 19, wherein the visual example is a visual example of a preposition.