Abstract: A data partitioning plan is automatically generated that—given a data-parallel program and a large input dataset, and without having to first run the program on the input dataset—substantially optimizes performance of the distributed execution system that explicitly measures and infers various properties of both data and computation to perform cost estimation and optimization. Estimation may comprise inferring the cost of a candidate data partitioning plan, and optimization may comprise generating an optimal partitioning plan based on the estimated costs of computation and input/output.

Abstract: Image search results are obtained by providing weights to visual features to emphasize features corresponding to objects of interest while simultaneously deemphasizing irrelevant or inconsistent features that lead to poor search results. In order to minimize the impact of visual features that are unreliable or irrelevant with respect to the objects of interest in the image, context-dependent weights are provided to detect visual features such that those visual features pertaining to the objects of interest are more heavily weighted than those visual features that pertain to irrelevant or unreliable portions of the image. Visual features may be weighted for images in a searchable database. Training data may be obtained and used in weighting visual features in a query image and, alternatively, in searchable database images.

Abstract: Text in web pages or other text documents may be classified based on the images or other objects within the webpage. A system for identifying and classifying text related to an object may identify one or more web pages containing the image or similar images, determine topics from the text of the document, and develop a set of training phrases for a classifier. The classifier may be trained and then used to analyze the text in the documents. The training set may include both positive examples and negative examples of text taken from the set of documents. A positive example may include captions or other elements directly associated with the object, while negative examples may include text taken from the documents, but from a large distance from the object. In some cases, the system may iterate on the classification process to refine the results.

Abstract: A Mixed Media Reality (MMR) system and associated techniques are disclosed. The MMR system provides mechanisms for forming a mixed media document that includes media of at least two types (e.g., printed paper as a first medium and digital content and/or web link as a second medium). The present invention provides a system, method, and computer program product for creating an ad hoc association between users. Captured digital images of paper documents or portions associated with a user are received. Document recognition features are extracted from the captured digital image. An ad hoc association is created between the user and another user associated with an existing document recognition feature similar to the extracted document recognition features.

Abstract: Product images are used in conjunction with textual descriptions to improve classifications of product offerings. By combining cues from both text and image descriptions associated with products, implementations enhance both the precision and recall of product description classifications within the context of web-based commerce search. Several implementations are directed to improving those areas where text-only approaches are most unreliable. For example, several implementations use image signals to complement text classifiers and improve overall product classification in situations where brief textual product descriptions use vocabulary that overlaps with multiple diverse categories. Other implementations are directed to using text and images “training sets” to improve automated classifiers including text-only classifiers.

Abstract: An image in a web page may be annotated after deriving information about an image when the image may be displayed on multiple web pages. The web pages that show the image may be analyzed in light of each other to determine metadata about the image, then various additional content may be added to the image. The additional content may be hyperlinks to other webpages. The additional content may be displayed as annotations on top of the images and in other manners. Many embodiments may perform searching, analysis, and classification of images prior to the web page being served.

Abstract: A set of data points is divided into a plurality of subsets of data points. A set of cluster closures is generated based at least in part on the subset of data points. Each cluster closure envelopes a corresponding cluster of a set of clusters and is comprised of data points of the enveloped cluster and data points neighboring the enveloped cluster. A k-Means approximator iteratively assigns data points to a cluster of the set of clusters and updates a set of cluster centroids corresponding to the set of clusters. The k-Means approximator assigns data points based at least in part on the set of cluster closures.

Abstract: Methods are disclosed for finding images from a large corpus of images that at least partially match a query image. The present method makes use of feature detectors to bundle features into local groups or bundles. These bundled features are repeatable and much more discriminative than an individual SIFT feature. Equally importantly, the bundled features provide a flexible representation that allows simple and robust geometric constraints to be efficiently enforced when querying the index.

Abstract: Product images are used in conjunction with textual descriptions to improve classifications of product offerings. By combining cues from both text and image descriptions associated with products, implementations enhance both the precision and recall of product description classifications within the context of web-based commerce search. Several implementations are directed to improving those areas where text-only approaches are most unreliable. For example, several implementations use image signals to complement text classifiers and improve overall product classification in situations where brief textual product descriptions use vocabulary that overlaps with multiple diverse categories. Other implementations are directed to using text and images “training sets” to improve automated classifiers including text-only classifiers.

Abstract: Trusted user accounts of an application provider are determined. Graphs, such as trees, are created with each node corresponding to a trusted account. Each of the nodes is associated with a vouching quota, or the nodes may share a vouching quota. Untrusted user accounts are determined. For each of these untrusted accounts, a trusted user account that has a social networking relationship is determined. If the node corresponding to the trusted user account has enough vouching quota to vouch for the untrusted user account, then the quota is debited, a node is added for the untrusted user account to the graph, and the untrusted user account is vouched for. If not, available vouching quota may be borrowed from other nodes in the graph.

Abstract: Text in web pages or other text documents may be classified based on the images or other objects within the webpage. A system for identifying and classifying text related to an object may identify one or more web pages containing the image or similar images, determine topics from the text of the document, and develop a set of training phrases for a classifier. The classifier may be trained and then used to analyze the text in the documents. The training set may include both positive examples and negative examples of text taken from the set of documents. A positive example may include captions or other elements directly associated with the object, while negative examples may include text taken from the documents, but from a large distance from the object. In some cases, the system may iterate on the classification process to refine the results.

Abstract: An efficient manner of performing an M-out-of-N partial matching search of indexed documents (e.g., web pages) is provided herein. More particularly, indexed words are arranged into a global location space (GLS), providing for respective occurrences of words in indexed documents being searched to have continuous locations on a one-dimensional GLS. Documents within the GLS are separated by end of document word marking boundaries between consecutive documents. The query words are then separated into an active set, comprising the left-most query words, and a non-active set. A partial matching operator transverses the GLS, applying active geometric constraints, in a sequential manner, to words in the active set. This causes shifting of the active set along the GLS to comprise M left-most query words. If a document satisfies constraints associated with M words in an active set, the document comprises at least M-out-of-N words.

Abstract: An image in a web page may be annotated after deriving information about an image when the image may be displayed on multiple web pages. The web pages that show the image may be analyzed in light of each other to determine metadata about the image, then various additional content may be added to the image. The additional content may be hyperlinks to other webpages. The additional content may be displayed as annotations on top of the images and in other manners. Many embodiments may perform searching, analysis, and classification of images prior to the web page being served.

Abstract: Text in web pages or other text documents may be classified based on the images or other objects within the webpage. A system for identifying and classifying text related to an object may identify one or more web pages containing the image or similar images, determine topics from the text of the document, and develop a set of training phrases for a classifier. The classifier may be trained and then used to analyze the text in the documents. The training set may include both positive examples and negative examples of text taken from the set of documents. A positive example may include captions or other elements directly associated with the object, while negative examples may include text taken from the documents, but from a large distance from the object. In some cases, the system may iterate on the classification process to refine the results.

Abstract: A system for identifying individuals in digital images and for providing matching digital images is provided. A set of images that include faces of known individuals is received. Faces are detected in the images and facial components are identified in each face. Visual words corresponding to the facial components are generated, stored, and associated with identifiers of the individuals. At a later time, a user may provide an image that includes the face of one of the known individuals. Visual words are determined from the face of the individual in the provided image and matched against the stored visual words. Images associated with matching visual words are ranked and presented to the user.

Abstract: Trusted user accounts of an application provider are determined. Graphs, such as trees, are created with each node corresponding to a trusted account. Each of the nodes is associated with a vouching quota, or the nodes may share a vouching quota. Untrusted user accounts are determined. For each of these untrusted accounts, a trusted user account that has a social networking relationship is determined. If the node corresponding to the trusted user account has enough vouching quota to vouch for the untrusted user account, then the quota is debited, a node is added for the untrusted user account to the graph, and the untrusted user account is vouched for. If not, available vouching quota may be borrowed from other nodes in the graph.

Abstract: A data partitioning plan is automatically generated that—given a data-parallel program and a large input dataset, and without having to first run the program on the input dataset—substantially optimizes performance of the distributed execution system that explicitly measures and infers various properties of both data and computation to perform cost estimation and optimization. Estimation may comprise inferring the cost of a candidate data partitioning plan, and optimization may comprise generating an optimal partitioning plan based on the estimated costs of computation and input/output.

Abstract: Images in a database or collection of images are each divided into multiple partitions with each partition corresponding to an area of an image. The partitions in an image may overlap with each other. Min-hash sketches are generated for each of the partitions and stored with the images. A user may submit an image and request that an image that is a partial match for the submitted image be located in the image collection. The submitted image is similarly divided into partitions and min-hash sketches are generated from the partitions. The min-hash sketches are compared with the stored min-hash sketches for matches, and images having partitions whose sketches are matches are returned as partial matching images.

Abstract: Methods and computer-readable media for propagating content category information to images stored in a database are described. A seed image that is associated with a known content category is received. A content-based image retrieval is conducted using the seed image as a search query image. A number of search result images are identified. The content category is propagated to the search result images. Metadata associated with the search result images is aggregated and analyzed to identify domains that should also be associated with the content category. Additional images that are associated with the domain are identified and the content category propagated thereto. The process is iterated using the additional images as search query images for the content-based image retrieval.

Abstract: Product images are used in conjunction with textual descriptions to improve classifications of product offerings. By combining cues from both text and image descriptions associated with products, implementations enhance both the precision and recall of product description classifications within the context of web-based commerce search. Several implementations are directed to improving those areas where text-only approaches are most unreliable. For example, several implementations use image signals to complement text classifiers and improve overall product classification in situations where brief textual product descriptions use vocabulary that overlaps with multiple diverse categories. Other implementations are directed to using text and images “training sets” to improve automated classifiers including text-only classifiers.