Uber Technologies, Inc.

Nokia Corporation

Percipient.ai Inc.

Nokia Technologies Oy

Vasudev Parameswaran

Radek Grzeszczuk

Jingchen Liu

Yanxi Liu

Varsha Hedau

Jerome Berclaz

Thommen Korah

Mohammed Waleed Kadous

Sheng Yang

Shivendra Pratap Singh

Chandan Prakash Sheth

Jane Alam Jan

Shang-Hsuan Tsai

Timo Pekka Pylvaenaeinen

Balan Rama Ayyar

Ivan Kovtun

Rajendra Jayantilal Shah

Vijay Mahadevan

Aditya Sarawgi

Ankit Tandon

Timo Pekka Pylvänäinen

Thommem Korah

Derek Young

Vidyadayini Talapady

Alison Higuera

Simon Chen

Atul Kanaujia

Reechik Chatterjee

Nikhil Kumar Gupta

Mojtaba Solgi

Anantha Krishnan Bangalore

This description describes a system for identifying individuals within a digital file. The system accesses a digital file describing the movement of unidentified individuals and detects a face for an unidentified individual at a plurality of locations in the video. The system divides the digital file into a set of segments and detects a face of an unidentified individual by applying a detection algorithm to each segment. For each detected face, the system applies a recognition algorithm to extract feature vectors representative of the identity of the detected faces which are stored in computer memory. The system applies a recognition algorithm to query the extracted feature vectors for target individuals by matching unidentified individuals to target individuals, determining a confidence level describing the likelihood that the match is correct, and generating a report to be presented to a user of the system.

Identification of individuals in a digital file using media analysis techniques

A computer vision system configured for detection and recognition of objects in video and still imagery in a live or historical setting uses a teacher-student object detector training approach to yield a merged student model capable of detecting all of the classes of objects any of the teacher models is trained to detect. Further, training is simplified by providing an iterative training process wherein a relatively small number of images is labeled manually as initial training data, after which an iterated model cooperates with a machine-assisted labeling process and an active learning process where detector model accuracy improves with each iteration, yielding improved computational efficiency. Further, synthetic data is generated by which an object of interest can be placed in a variety of setting sufficient to permit training of models. A user interface guides the operator in the construction of a custom model capable of detecting a new object.

Systems and methods for rapid development of object detector models

An authoritative candidate is selected for determining a location of a point of interest (POI). Source data including name, address, and location for POIs is received from multiple data sources. The received data is normalized for ease of comparison, and coordinates for each candidate are compared to coordinates of other candidates to determine which candidate if any is an authoritative location for the POI. The candidate locations are compared using two models a metric-based scoring system and a machine learning model that may utilize a gradient boosted decision tree. The authoritative candidate can be used to render digital maps that include the POI. In addition, the authoritative candidate's location can be used to provide vehicle route guidance to the POI.

Vehicle routing guidance to an authoritative location for a point of interest

A system trains and applies a machine learning model to label maps of a region. Various data modalities are combined as inputs for multiple data tiles used to characterize a region for a geographical map. Each data modality reflects sensor data captured in different ways. Some data modalities include aerial imagery, point cloud data, and location trace data. The different data modalities are captured independently and then aggregated using machine learning models to determine map labeling information about tiles in the region. Data is ingested by the system and corresponding tiles are identified. A tile is represented by a feature vector of different data types related to the various data modalities, and values from the ingested data are added to the feature vector for the tile. Models can be trained to predict characteristics of a region using these various types of input.

Determining geographical map features with multi-sensor input

Architecture that detects entrances on building facades. In a first stage, scene geometry is exploited and the multi-dimensional problem is reduced down to a one-dimensional (1D) problem. Entrance hypotheses are generated by considering pairs of locations along lines exhibiting strong gradients in the transverse direction. In a second stage, a rich set of discriminative image features for entrances is explored according to constructed designs, specifically focusing on properties such as symmetry and color consistency, for example. Classifiers (e.g., random forest) are utilized to perform automatic feature selection and entrance classification. In another stage, a joint model is formulated in three dimensions (3D) for entrances on a given facade, which enables the exploitation of physical constraints between different entrances on the same facade in a systematic manner to prune false positives, and thereby select an optimum set of entrances on a given facade.

Entrance detection from street-level imagery

A map labeling system trains a machine-learned model using a set of training data and generates a set of test predictions for a set of test properties by applying the machine-learned model to a set of testing data. Each prediction in the set of test predictions comprises a confidence score representing the machine-learned model's confidence in the prediction. The map labeling system determines a correctness of each prediction in the set of predictions and determines a relationship between the confidence scores and the correctness of the test predictions. The map labeling system establishes a confidence threshold for the machine-learned model based on the determined relationship and labels a production property by applying the machine-learned model to production data.

Confidence threshold determination for machine-learned labeling

A location detection system performs a process that uses data from a satellite navigation sensor in conjunction with data from a second sensor to determine the accuracy of location estimates provided by the satellite navigation sensor. The system uses the satellite navigation sensor to determine location estimates from at a first time and a second time. The system also uses data from the second sensor to determine a third location estimate, which represents another estimate of the system's location at the second time. The system uses the three location estimates to determine whether the second location estimate satisfies an accuracy condition. If the accuracy condition is satisfied, then the second location estimate may be provided as input to a process.

Location accuracy determination based on data from second sensor

A system adjusts an estimated travel time from an origin to a destination to better predict an actual trip duration. The system receives a route from a specified origin and destination. The system receives an estimated trip duration corresponding to the generated route. A machine learned model improves the estimated trip duration by applying data about past trips facilitated by the system and data about traffic-control features associated with a route. For example, the system may use counts of a number of road signs and a number of traffic signals located along the generated route to predict an actual trip duration. In some cases, the system may additionally use data about synchronized traffic lights to predict actual trip duration.

Determining adjusted trip duration using route features

A digital image is converted to a multiple level image, and multiple scale sets are formed from connected components of the multiple level image such that different ones of the scale sets define different size spatial bins. For each of the multiple scale sets there is generated a count of connected components extracted from the respective scale set for each spatial bin; and adjacent spatial bins which represent connected components are linked. Then the connected components from the different scale sets are merged and text line detection is performed on the merged connected components. In one embodiment each of the scale sets is a histogram, and prior to linking all bins with less than a predetermined count are filtered out; and each histogram is extended such that counts of adjacent horizontal and vertical bins are added (single region bins are filtered out) and the linking is on the extended histograms.

Text detection using multi-layer connected components with histograms

Methods and apparatuses are provided for facilitating detection of text within an image. A method may include calculating an alpha value associated with an image region containing a hypothesized text fragment. The alpha value may be defined as a function of a curved character length distribution, a character width distribution, and an inter-character spacing distribution for the hypothesized text fragment. The method may additionally include calculating a gamma value based at least in part on an interval length distribution determined for the hypothesized text fragment. The method may also include classifying whether the image region is a text-containing region based at least in part on the calculated alpha and gamma values. Corresponding apparatuses are also provided.

Methods and apparatuses for facilitating detection of text within an image

Growing community of inventors

Fremont, CA, United States of America

Vasudev Parameswaran