Abstract: Methods for detecting a horizon in an image with a plurality of pixels can include the step of blurring the image with a noise filter, then dividing the image into an M×N matrix of sub-blocks S. For each sub-block S, horizon features can be coarse-extracted by defining an r-dimensional vector having P feature values for each sub-block S and clustering each r-dimensional vectors into two clusters using a k-means statistical analysis. The corresponding sub-blocks S corresponding to the two clusters can be masked with a binary mask. The methods can further include the step of fine-extracting the horizon features at a pixel level for each sub-block Si, j and sub-block Si?1, j when the binary mask changes value from sub-block Si?1 j to said sub-block Si, j, for i=1 to M and j=1 to N.

Abstract: A method and system are provided for processing maritime video data to improve processing time and reduce the risk of false detection. Video data input may be received from satellites, aircrafts, UAVs and other aerial vehicles. The method and system create annotated video output files indicating the location of tracked anomalies. The method and system perform an anomaly detection function on each data frame within the video input, identify anomalous regions within each data frame, and identify clusters. The method and system then perform an algorithm to discard clusters which do not meet threshold criteria to reduce the risk of false detection, as well as employ a horizon detection algorithm to eliminate superfluous image data unrelated to tracking maritime targets.

Abstract: Techniques for context-based queryless presentation of recommendations are described. Without receiving an explicit query, an inferred query is determined based on a first set of context signals associated with a recommendation triggering event. Items matching the inferred query are selected based on a second set of one or more context signals associated with the recommendation triggering event. One or more results items identifying the one or more matching items are caused to display. A result items may be displayed in association with an explanation that identifies context values used to select an item for recommendation or one or more attributes of the item that are relevant to a user or the inferred query.