Microsoft Technology Licensing, LLC

Maksim Lepikhin

David J Simons

Dragomir Dimitrov Yankov

James Evan Undery

Wolfgang Schickler

Christopher Stephen Messer

Leon Rosenshein

Michael Robert Evans

Philip Starhill

Chiqun Zhang

Chintan Anil Shah

Corey Benjamin Goff, Jr

Keith Seifert

Technology is described herein for ranking candidate result items in at least two stages. In a first stage, the technology uses a first attention-based neural network to determine an extent of attention that each token of an input query should pay to the tokens of each candidate result item. In a second stage, the technology uses a ranking subsystem to perform listwise inference on output results provided by the first stage, to generate a plurality of ranking scores that establish an order of relevance of the candidate results items. The ranking subsystem may use a second attention-based neural network to perform the listwise inference. According to some implementations, the technology is configured to process queries and candidate result items having different kinds and combinations of features. For instance, one kind of input query may include text-based features, structure-based features, and geographic-based features.

Processing queries using an attention-based ranking system

An input mesh can be decomposed into component meshes that can be independently simplified. A computing device can calculate costs of performing candidate edge collapses for a component mesh. The candidate edge collapses can include boundary edge collapses and interior edge collapses. To simplify a component mesh, the execution of boundary edge collapses and the execution of interior edge collapses are interleaved in an order based on the costs of performing the candidate edge collapses. The position of a vertex resulting from a boundary edge collapse can be calculated independently of the interior of the component mesh. When component meshes are simplified in parallel, a boundary that is common to the component meshes can be simplified identically.

Maintaining consistent boundaries in parallel mesh simplification

Many computing services incorporate large mosaics of geo-located images. For example, a user may access and interact with a digital image map comprising thousands of digital images mosaiced together. It may be advantageous to parallelize mosaic creation by splitting the mosaic into tiles, and processing the tiles independently from one another. Unfortunately, visual artifacts between image boundaries may result because of the independent processing. Accordingly, one or more systems and/or techniques for generating a substantially seamless mosaic tile are disclosed herein. A mosaic tile comprising one or more intersecting images may be received. Pair-wise image segmentation may be executed upon pairs of intersecting images within the mosaic tile to generate seam lines between the pair of intersecting images. The seam lines are used to compose the mosaic tile. A constrained alpha-expansion algorithm is executed upon hole regions within the composed mosaic to generate a substantially seamless mosaic.

Parallelized generation of substantially seamless image mosaics

Among other things, one or more techniques and/or systems are provided for quantifying blur of an image. Blur may result due to motion of a camera while the image is captured. Accordingly, motion measurement data corresponding to motion of the camera during an exposure event may be used to create a camera rotation matrix. A camera intrinsic matrix may be obtained based upon a focal length and principle point of the camera. A transformation matrix may be estimated based upon the camera rotation matrix and/or the camera intrinsic matrix. The transformation matrix may be applied to pixels within the image to determine a blur metric for the image. In this way, blur of an image may be quantified offline and/or in real-time during operation of the camera (e.g., so that the image may be re-acquired (e.g., on the fly) if the image is regarded as being overly blurry).

Image blur detection

When photographs are to be combined into a single image, haze correction and/or color balancing may be performed. The photographs may be analyzed and left-clipped in order to darken the photographs and to increase the density of pixels in the low-luminosity region, thereby decreasing the perception of haze. When the photographs are combined into one continuous image, tie points are selected that lie in regions where the photographs overlap. The tie points may be selected based on visual similarity of the photographs in the region around the tie point, using a variety of algorithms. Functions are then chosen to generate saturation and luminosity values that minimize, at the tie points, the cost of using the generated values as opposed to the actual saturation and luminosity values. These functions are then used to generate saturation and luminosity values for the full image.

Growing community of inventors

Boulder, CO, United States of America

Maksim Lepikhin