Abstract: Techniques for determining motion saliency in video content using center-surround receptive fields. In some implementations, images or frames from a video may be apportioned into non-overlapped regions, for example, by applying a rectilinear grid. For each grid region, or cell, motion consistency may be measured between the center and surround area of that cell across frames of the video. Consistent motion across the center-surround area may indicate that the corresponding region has low variation. The larger the difference between center-surround motions in a cell, the more likely the region has high motion saliency.

Abstract: The disclosed technology includes techniques for improved content coverage in automatically-generated content summaries. The technique may include clustering a set of input content, determining diffusion for each cluster, and selecting representatives of each cluster to optimize other secondary metrics. Various types of input content may be used, including groups of images, video clips, or other multimedia content. Contiguous content may be manually or programmatically divided into discrete portions before clustering, for example, a lengthy video divided into a number of short clips. In some implementations, the disclosed technique may be implemented effectively on a mobile device. In other words, the processing required may be computationally feasible for execution on a smartphone or similar device.

Abstract: A method and apparatus for image processing a lens-distorted image (e.g., a fisheye image) is provided. The method includes partitioning coordinate points in a selected output image into tiles. The output image is an undistorted rendition of a subset of the lens-distorted image. Coordinate points on a border of the tiles in the output image are selected. For each tile, coordinate points in the lens-distorted image corresponding to each selected coordinate point in the output image are calculated. In addition, for each tile, a bounding box on the lens-distorted image is selected. The bounding box includes the calculated coordinates in the lens-distorted image. The bounding boxes are expanded so that they encompass all coordinate points in the lens-distorted image that map to all coordinate points in their respective corresponding tiles. Output pixel values are generated for each tile from pixel values in their corresponding expanded bounding boxes.

Abstract: Certain embodiments of the disclosed technology include systems and methods for determining the priority of a notification on a mobile device using machine learning. Other aspects of the disclosed technology include selectively displaying notifications based on the priority of a notification.

Abstract: Certain embodiments of the disclosed technology include systems and methods for determining the priority of a notification on a mobile device using machine learning. Other aspects of the disclosed technology include selectively displaying notifications based on the priority of a notification.

Abstract: A video hosting service comprising video classifiers that identify content sources of content included in videos uploaded to the video hosting service. Identifying the content source allows a content owner of the content source to claim ownership of videos that include content based on the content source. Usage policies associated with the content owners are applied to the uploaded videos that describe how the video hosting service is to treat the videos.

Abstract: Methods, systems and articles of manufacture for identifying semantic nearest neighbors in a feature space are described herein. A method embodiment includes generating an affinity matrix for objects in a given feature space, wherein the affinity matrix identifies the semantic similarity between each pair of objects in the feature space, training a multi-bit hash function using a greedy algorithm that increases the Hamming distance between dissimilar objects in the feature space while minimizing the Hamming distance between similar objects, and identifying semantic nearest neighbors for an object in a second feature space using the multi-bit hash function. A system embodiment includes a hash generator configured to generate the affinity matrix and train the multi-bit hash function, and a similarity determiner configured to identify semantic nearest neighbors for an object in a second feature space using the multi-bit hash function.

Abstract: Disclosed are apparatus and methods for providing machine-learning services. A context-identification system executing on a mobile platform can receive data comprising context-related data associated with the mobile platform and application-related data received from the mobile platform. The context-identification system can identify a context using the context-related data associated with the mobile platform and/or the application-related data received from the mobile platform. Based on at least one context identified, context-identification system can predict a communicative action associated with the mobile platform by performing a machine-learning operation on the received data. An instruction can be received to execute the communicative action associated with the mobile platform.

Abstract: Methods and apparatus for predicting time spans for mobile platform activation are presented. A machine-learning service executing on a mobile platform receives feature-related data. The feature-related data includes usage-related data about time spans that the mobile platform is activated and platform-related data received from the mobile platform. The usage-related data and the platform-related data can differ. The machine-learning service determines whether the machine-learning service is trained to perform machine-learning operations related to predicting a time span that the mobile platform will be activated. In response to determining that the machine-learning service is trained, the machine-learning service: receives a request for a predicted time span that the mobile platform will be activated, determines the predicted time span by the machine-learning service performing a machine-learning operation on the feature-related data, and sends the predicted time span.

Abstract: Disclosed are apparatus and methods for providing machine-learning services. A machine-learning service executing on a mobile platform can receive data related to a plurality of features. In some cases, the received data can include data related to features received from an application and data related to features received from the mobile platform. The machine-learning service can determine at least one feature based on the received data. The machine-learning service can generate an output by performing a machine-learning operation on the at least one feature. The machine-learning operation can be selected from among an operation of ranking the at least one feature, an operation of classifying the at least one feature, an operation of predicting the at least one feature, and an operation of clustering the at least one feature. The machine-learning service can send the output.

Abstract: A system and an apparatus for improving image quality of an object in video telephony are described. The apparatus comprises an image detection unit, an image alignment unit and an image fusion unit. A near infrared light source in the image detection unit illuminates the object, such that the object is front illuminated. An image sensor alongside the near infrared light source generates a near infrared image and a visible image. An image alignment unit aligns the near infrared image and the visible image. An image fusion unit fuses the aligned near infrared image and aligned visible image pair to form a composite image of the object. The composite image of the object has improved image quality, image detail, and a reduction in shadows.

Abstract: A method and apparatus for image processing a lens-distorted image (e.g., a fisheye image) is provided. The method includes partitioning coordinate points in a selected output image into tiles. The output image is an undistorted rendition of a subset of the lens-distorted image. Coordinate points on a border of the tiles in the output image are selected. For each tile, coordinate points in the lens-distorted image corresponding to each selected coordinate point in the output image are calculated. In addition, for each tile, a bounding box on the lens-distorted image is selected. The bounding box includes the calculated coordinates in the lens-distorted image. The bounding boxes are expanded so that they encompass all coordinate points in the lens-distorted image that map to all coordinate points in their respective corresponding tiles. Output pixel values are generated for each tile from pixel values in their corresponding expanded bounding boxes.

Abstract: A method and apparatus for increasing resolution of video images are disclosed. Vectors may be produced based on a sequence of video frames. Low dimensional vectors may be produced from the vectors. Groups of at least two of the low dimensional vectors may be interpolated to produce respective low dimensional interpolated vectors. Each of the low dimensional vectors and the interpolated low dimensional vectors may be mapped, according to a model, to obtain dimensionally increased image information. Aspects of the image information may be included in corresponding video frames and corresponding interpolated video frames.

Abstract: Straight lines in a compensated image of a wide-angle image are identified (102) and corresponding curved lines of pixels in the wide-angle image identified (103) as well. For given points in the aforementioned straight lines, a corresponding point in the corresponding curved lines is identified by determining an intersection of those curved lines with another line of pixels in the wide-angle image (104). The latter can comprise either a curved line that corresponds to another straight line in the compensated image, in which case the intersection is noted by tracing the curved lines using a scan-rasterization algorithm. The latter can also comprise, if desired, a straight line that corresponds to another straight line in the compensated image. In either case, the point of intersection is readily and efficiently determined.

Abstract: A method and apparatus for image processing a lens-distorted image (e.g., a fisheye image) is provided. The method includes partitioning coordinate points in a selected output image into tiles. The output image is an undistorted rendition of a subset of the lens-distorted image. Coordinate points on a border of the tiles in the output image are selected. For each tile, coordinate points in the lens-distorted image corresponding to each selected coordinate point in the output image are calculated. In addition, for each tile, a bounding box on the lens-distorted image is selected. The bounding box includes the calculated coordinates in the lens-distorted image. The bounding boxes are expanded so that they encompass all coordinate points in the lens-distorted image that map to all coordinate points in their respective corresponding tiles. Output pixel values are generated for each tile from pixel values in their corresponding expanded bounding boxes.

Abstract: Visual tracking over a sequence of images is formulated by defining an object class and one or more background classes. The most discriminant features available in the images are then used to select a portion of each image as belonging to the object class. Fisher's linear discriminant method is used to project high-dimensional image data onto a lower-dimensional space, e.g., a line, and perform classification in the lower-dimensional space. The projection function is incrementally updated.

Abstract: Straight lines in a compensated image of a wide-angle image are identified (102) and corresponding curved lines of pixels in the wide-angle image identified (103) as well. For given points in the aforementioned straight lines, a corresponding point in the corresponding curved lines is identified by determining an intersection of those curved lines with another line of pixels in the wide-angle image (104). The latter can comprise either a curved line that corresponds to another straight line in the compensated image, in which case the intersection is noted by tracing the curved lines using a scan-rasterization algorithm. The latter can also comprise, if desired, a straight line that corresponds to another straight line in the compensated image. In either case, the point of intersection is readily and efficiently determined.

Abstract: A method and apparatus for increasing resolution of video images are disclosed. Vectors may be produced based on a sequence of video frames. Low dimensional vectors may be produced from the vectors. Groups of at least two of the low dimensional vectors may be interpolated to produce respective low dimensional interpolated vectors. Each of the low dimensional vectors and the interpolated low dimensional vectors may be mapped, according to a model, to obtain dimensionally increased image information. Aspects of the image information may be included in corresponding video frames and corresponding interpolated video frames.

Abstract: A system and an apparatus for improving image quality of an object in video telephony are described. The apparatus comprises an image detection unit, an image alignment unit and an image fusion unit. A near infrared light source in the image detection unit illuminates the object, such that the object is front illuminated. An image sensor alongside the near infrared light source generates a near infrared image and a visible image. An image alignment unit aligns the near infrared image and the visible image. An image fusion unit fuses the aligned near infrared image and aligned visible image pair to form a composite image of the object. The composite image of the object has improved image quality, image detail, and a reduction in shadows.

Abstract: A system and a method are disclosed for adaptive probabilistic tracking of an object within a motion video. The method utilizes a time-varying Eigenbasis and dynamic, observation and inference models. The Eigenbasis serves as a model of the target object. The dynamic model represents the motion of the object and defines possible locations of the target based upon previous locations. The observation model provides a measure of the distance of an observation of the object relative to the current Eigenbasis. The inference model predicts the most likely location of the object based upon past and present observations. The method is effective with or without training samples. A computer-based system provides a means for implementing the method. The effectiveness of the system and method are demonstrated through simulation.