Abstract: A feature extraction system extracts map features from an aerial image. The feature extraction system receives an aerial image having pixels and predicts, for each pixel, a probability that the pixel corresponds to a map feature based on a machine learning model. The machine learning model is trained to determine a probability that a pixel corresponds to the map feature based on a training dataset comprising pairs of aerial images and corresponding mask images that describe known instances of the map feature. The feature extraction system identifies a subset of pixels of the plurality of pixels. Each pixel in the subset has a predicted probability that is greater than or equal to a threshold probability that a pixel corresponds to the map feature. The feature extraction system further determines a bounded geometry enclosing the identified subset of pixels, the bounding geometry encompassing an instance of the map feature.

Abstract: A feature extraction system extracts map features from an aerial image. The feature extraction system receives an aerial image having pixels and predicts, for each pixel, a probability that the pixel corresponds to a map feature based on a machine learning model. The machine learning model is trained to determine a probability that a pixel corresponds to the map feature based on a training dataset comprising pairs of aerial images and corresponding mask images that describe known instances of the map feature. The feature extraction system identifies a subset of pixels of the plurality of pixels. Each pixel in the subset has a predicted probability that is greater than or equal to a threshold probability that a pixel corresponds to the map feature. The feature extraction system further determines a bounded geometry enclosing the identified subset of pixels, the bounding geometry encompassing an instance of the map feature.

Abstract: A feature extraction system extracts map features from an aerial image. The feature extraction system receives an aerial image having pixels and predicts, for each pixel, a probability that the pixel corresponds to a map feature based on a machine learning model. The machine learning model is trained to determine a probability that a pixel corresponds to the map feature based on a training dataset comprising pairs of aerial images and corresponding mask images that describe known instances of the map feature. The feature extraction system identifies a subset of pixels of the plurality of pixels. Each pixel in the subset has a predicted probability that is greater than or equal to a threshold probability that a pixel corresponds to the map feature. The feature extraction system further determines a bounded geometry enclosing the identified subset of pixels, the bounding geometry encompassing an instance of the map feature.

Abstract: A feature extraction system extracts map features from an aerial image. The feature extraction system receives an aerial image having pixels and predicts, for each pixel, a probability that the pixel corresponds to a map feature based on a machine learning model. The machine learning model is trained to determine a probability that a pixel corresponds to the map feature based on a training dataset comprising pairs of aerial images and corresponding mask images that describe known instances of the map feature. The feature extraction system identifies a subset of pixels of the plurality of pixels. Each pixel in the subset has a predicted probability that is greater than or equal to a threshold probability that a pixel corresponds to the map feature. The feature extraction system further determines a bounded geometry enclosing the identified subset of pixels, the bounding geometry encompassing an instance of the map feature.

Abstract: A method for identifying missing map features in an electronic map, involving receiving telemetry probes indicating a geographic location of a mobile computing device, and identifying a subset of telemetry probes corresponding to an existing map feature. The identified subset is then removed from an aggregation of telemetry probes, and the remaining telemetry probes used to generate a density map and identify missing clusters of telemetry probes. A geometry of the missing clusters is determined, and a missing map feature defined from the geometry of the missing cluster. An electronic map may be updated with the missing map feature.