The present invention relates generally to vision systems, and more particularly, to a method and apparatus for searching videos based on a mapping from a set of physical concepts to visual properties or descriptors, without requiring the user to know the underlying properties and their values, or perform the translation manually.
Database searching tools exist for all sorts of queries, including video. When an user is searching for objects in video clips stored in a video database, the most natural query consists of nouns representing concepts such as, for example, “person,” “vehicle,” “convoy,” or “building.” Similarly, activities are represented by combinations of nouns and verbs such as “vehicle”/“turn.” This is the model followed by some popular video search tools, such as Google Video™. In a Google Video™ keyword search, the search term(s) need to match a caption/annotation associated with a video clip in a video database. Once again, vocabulary mismatch presents a key challenge if the user query must be compared against video annotations. For example, if the video is not annotated with the same keywords, then no result will be returned.
Retrieval performance may be improved over the method of searching with simple keyword search terms that are matched to video annotations. One method that is well-documented in the information retrieval literature is known as query expansion. In the text retrieval domain, a number of highly-ranked documents (i.e., document content) are reissued as a new query, thereby expanding the query with additional query terms. In the video retrieval domain, there is also a body of computer vision literature devoted to query expansion. In “Total Recall: Automatic Query Expansion with a Generative Feature Model for Object Retrieval,” (0. Chum, J. Philbin, J. Sivic, M. Isard, and A. Zisserman), ICME, 2007, (Chum et al.), a bag-of-visual-words architecture is adopted to achieve high precision. Chum et al. also presents two contributions to query expansion—the use of strong spatial constraints between image and each result and learning of a latent feature model from the images. The drawback of the approach of Chum et al. is that feature detection and quantization are noisy processes, leading to variation in the visual words and consequently missed results. In “Semantic Concept-Based Query Expansion and Re-ranking for Multimedia Retrieval,” (A. Natsev, A. Haubold, J. Tesic, L. Xi, and R. Yan), ACM Multimedia, 2007, (Natsev et al.), approaches for query expansion are presented in which textual keywords, visual examples or initial retrieval results are analyzed to identify the most relevant visual concepts for a given query. The approaches of Natsev et al. are both lexical and involve statistical corpus analysis, which require deep parsing or semantic tagging of queries or lexical query expansion. In “Enabling Video Annotation Using a Semantic Database Extended with Visual Knowledge,” (G. Stein, J. Rittscher, A. Hoogs), ICME, 2003, (“Stein et al.”), an extension to WordNet is described that contains specific visual information (WordNet is a semantic lexicon for the English language. It groups English words into sets of synonyms called synsets, provides short, general definitions, and records the various semantic relations between these synonym sets. WordNet was created and has been maintained at the Cognitive Science Laboratory of Princeton University, Princeton, N.J.). However, the Stein et al. paper focuses on how such a semantic database makes video annotation possible for Broadcast News. In “Creating a Geospatial and Visual Information Ontology for Users,” (C. Basu, H. Cheng, C. Fellbaum), Ontology for the Intelligence Community, 2007 (“the GVIO paper”), which is incorporated herein by reference in its entirety, an extension to WordNet is also developed. The focus of the GVIO paper is on a different aspect of query expansion than is presented in the present invention.
Accordingly, what would be desirable, but has not yet been provided, is a system and method for effectively and automatically searching for, identifying, and retrieving high precision video clips from a database based on a mapping of a set of physical concepts to visual properties or descriptors.
The above-described problems are addressed and a technical solution is achieved in the art by providing a computer implemented method for retrieving video clips from a database, comprising the steps of retrieving in an initial query from a video collection based on a search term; receiving a user selection of at least one video clip from a first set of video clips corresponding to the search term; associating at least one visual attribute of the selected video clip with the search term; receiving the at least one search term from a user in a subsequent query; determining a set of physical concepts based on the at least one search term; mapping the set of physical concepts to a plurality of visual attributes; searching the database for at least one video clip corresponding to the plurality of visual attributes; identifying at least one video clip in the database having the plurality of visual attributes; and returning a second set of video clips having the set of visual attributes to the user, the second set including the at least one video clip. The second set may contain fewer video clips than the first set. According to an embodiment of the present invention, determining a set of physical concepts and mapping the set of physical concepts may be performed using a taxonomy and an inference engine. Determining a set of physical concepts may further comprise finding synonyms of the search term for use in determining the set of physical concepts. The method may further comprise the step of querying a plurality of collections of video clips in the database, wherein the range of values for a given visual attribute is the union of values that covers substantially all video clips having said given visual attribute across the plurality of collections of video clips. At least one of the plurality of visual attributes may be derived from sensor metadata stored with at least one of the second set of video clips. At least one of the plurality of visual attributes may be associated with the selected video clip.
According to an embodiment of the present invention, the method may further comprise the steps of extracting at least one actual value of at least one of the plurality of visual attributes for which at least one default value has been assigned in the taxonomy; associating with the at least one actual value at least one other visual attribute from the second set of video clips; and annotating the taxonomy with the at least one other associated visual attribute when available from a user selected video clip. The retrieval method may further comprise the steps of receiving the at least one search term from the user; determining a second set of physical concepts based on the at least one search term; mapping the second set of physical concepts to a second plurality of visual attributes based on the annotated taxonomy; searching the database for at least one video corresponding to the second plurality of visual attributes; identifying at least one video clip in the database having the second plurality of visual attributes; and returning a third set of video clips having the second plurality of visual attributes to the user, the third set including the at least one video clip. The third set may contain fewer video clips than the second set.
Default values may be assigned to the plurality of visual attributes, the default value being computed based on a collection of training video clips. Minimum and maximum values of visual attributes in the plurality of visual attributes may be pre-computed. A value corresponding to each of the plurality of visual attributes may be derived from metadata contained within a collection of training video clips.
The present invention will be more readily understood from the detailed description of exemplary embodiments presented below considered in conjunction with the attached drawings, of which:
It is to be understood that the attached drawings are for purposes of illustrating the concepts of the invention and may not be to scale.
Referring now to
Instead of retrieving videos in response to a query of a concept term based on external annotations of text, embodiments of the present invention reformulate or transform the query from keywords to a representative set of visual descriptors (properties) and their associated values, thereby harnessing a representation of visual information in sensor metadata stored with the video (Raw sensor metadata is data available as part of the actual video itself. Examples include geo-coordinates, time-of-day, and manual annotation. Other attributes may be derived or computed from sensor metadata stored with the video.). As a result, mappings between semantic information (i.e., concepts) and the sensor metadata are established.
Referring now to
In some embodiments, when one or more visual attributes are associated with a video clip during an initial query, the values of the visual attributes may be derived or calculated from one or more images in a collection of training video clips, which may or may not contain one or more of the video clips or video clip collection(s) in the database being queried at steps 32, 34. For a specific data collection, the attribute values computed over an aggregate of instances are referred to default values. Default values of visual attributes may be stored as slot-fillers in a knowledge base, such as a Protégé, application-specific knowledge base. Minimum and maximum values of default visual attributes may be pre-computed. A value corresponding to each of the visual attributes may be derived from metadata contained within a collection of training video clips. Using a rule-based inference system, such as Algernon, the Protégé knowledge base is queried and values are retrieved that have been pre-computed for a training video collection.
Referring now to
In other embodiments, the visual attributes may be derived directly from sensor metadata associated with selected video clip(s), or a combination of selected video clips and default values. In embodiments of the present invention, the values of visual attributes from selected video clips may replace one or more of the default values of the visual attributes in subsequent queries involving the same search term as previously entered. The selection of visual attribute values from current or prior selected video clips, or from previously calculated default values will be discussed in more detail hereinbelow.
Referring again to
When querying multiple video collections in the database, the range of values for a visual attribute may be the union of values that covers substantially all video clips having the visual attribute across the plurality of collections of video clips. The maximal set of values that cover all positive examples of video clips across the collections to form a search query is taken. For example, in collection 1, if the range of “slant angle” for “vehicle” is a subset of the range of “slant angle” in collection 2, then the two ranges are combined by taking the smallest maximal range that covers the possible values of “slant angle” of vehicles in collections 1 and 2 at query time. This may produce high recall at the expense of precision for the resulting set of retrieved video clips. In other words, all video clips that satisfy the “slant angle” constraint may be retrieved from both collections.
To increase the precision of the resulting set of retrieved video clips, query expansion can be extended by augmenting the mapping step 42 of
The results of step 74 (i.e., associating the values of visual attributes and video clips selected for a concept) may be available for the expansion (or generation) of queries in future searches. For example, during the next search session using the same search term, the user may be presented with a choice of previously selected video clips and associated property values as well as the original search screen populated with default values. An embodiment of query expansion in subsequent searches of the same concept is illustrated in the flow of
At step 76, the same search term previously entered in an initial query is received by the system. At step 78, a second set of physical concepts based on the at least one search term is determined. This second set of physical concepts is derived from the expansion of concepts determined in step 72 of
It is to be understood that the exemplary embodiments are merely illustrative of the invention and that many variations of the above-described embodiments may be devised by one skilled in the art without departing from the scope of the invention. It is therefore intended that all such variations be included within the scope of the following claims and their equivalents.
This application claims the benefit of U.S. provisional patent application No. 61/013,192 filed Dec. 12, 2007, the disclosure of which is incorporated herein by reference in its entirety.
This invention was made with U.S. government support under contract number NBCHC070062. The U.S. government has certain rights in this invention.
Number | Date | Country | |
---|---|---|---|
61013192 | Dec 2007 | US |