Waymo LLC

A9.com, Inc.

Edward Hsiao

Douglas Ryan Gray

Maoqing Yao

Abhijit S Ogale

Andreas Wendel

Yu Ouyang

David Yonchar Margines

Alexander L Honda

Maxim Krivokon

Yosuke Higashi

David Harrison Silver

Xiaofan Lin

Mark Jay Nitzberg

Nityananda Jayadevaprakash

Carl Kershaw

Jonathan Hsiao

Shruti Sheorey

Pok Man Chu

Aspects of the disclosure relate to controlling a vehicle having an autonomous driving mode. For instance, a current state of a traffic light may be determined. One of a plurality of yellow light durations may be selected based on the current state of the traffic light. When the traffic light will turn red may be predicted based on the selected one. The prediction may be used to control the vehicle in the autonomous driving mode.

Yellow light durations for autonomous vehicles

The technology relates to controlling a vehicle based on a railroad light's activation status. In one example, one or more processors receive images of a railroad light. The one or more processors determine, based on the images of the railroad light, the illumination status of a pair of lights of the railroad light over a period of time as the vehicle approaches the railroad light. The one or more processors determine based on the illumination status of the pair of lights, a confidence level, wherein the confidence level indicates the likelihood the railroad light is active. The vehicle is controlled as it approaches the railroad light based on the confidence level.

Railroad light detection

Aspects of the disclosure relate to training a machine learning model on a distributed computing system. The model can be trained using selected processors of the training platform. The distributed system automatically modifies the model for instantiation on each processor, adjusts an input pipeline to accommodate the capabilities of selected processors, and coordinates the training between those processors. Simultaneous processing at each stage can be scaled to reduce or eliminate bottlenecks in the distributed system. In addition, autonomous monitoring and re-allocating of resources can further reduce or eliminate bottlenecks. The training results may be aggregated by the distributed system, and a final model may then be transmitted to a user device.

Machine learning training platform

Aspects of the disclosure relate to detecting and responding to malfunctioning traffic signals for a vehicle having an autonomous driving mode. For instance, information identifying a detected state of a traffic signal for an intersection. An anomaly for the traffic signal may be detected based on the detected state and prestored information about expected states of the traffic signal. The vehicle may be controlled in the autonomous driving mode based on the detected anomaly.

Detecting and responding to malfunctioning traffic signals for autonomous vehicles

Machine-learning models are described detecting the signaling state of a traffic signaling unit. A system can obtain an image of the traffic signaling unit, and select a model of the traffic signaling unit that identifies a position of each traffic lighting element on the unit. First and second neural network inputs are processed with a neural network to generate an estimated signaling state of the traffic signaling unit. The first neural network input can represent the image of the traffic signaling unit, and the second neural network input can represent the model of the traffic signaling unit. Using the estimated signaling state of the traffic signaling unit, the system can inform a driving decision of a vehicle.

Detecting traffic signaling states with neural networks

Methods and system are provided for training and using a model to determine states of lanes of interest. For instance, image data including an image and an associated label identifying at least one traffic light, a state of the at least one traffic light, and a lane controlled by the at least one traffic light are received and used to train the model such that the model is configured to, in response to receiving an image and a lane of interest included in the image, output a lane state for the lane of interest. This model is then used by a vehicle in order to determine a state of a lane of interest. This state is then used to control the vehicle in an autonomous driving mode based on the state of the lane of interest.

Traffic light detection and lane state recognition for autonomous vehicles

An image creation and editing tool can use the data produced from training a neural network to add stylized representations of an object to an image. An object classification will correspond to an object representation, and pixel values for the object representation can be added to, or blended with, the pixel values of an image in order to add a visualization of a type of object to the image. Such an approach can be used to add stylized representations of objects to existing images or create new images based on those representations. The visualizations can be used to create patterns and textures as well, as may be used to paint or fill various regions of an image. Such patterns can enable regions to be filled where image data has been deleted, such as to remove an undesired object, in a way that appears natural for the contents of the image.

Neural network-based image manipulation

An object recognition system can be adapted to recognize subject matter having very few features or limited or no texture. A feature-sparse or texture-limited object can be recognized by complementing local features and/or texture features with color, region-based, shape-based, three-dimensional (3D), global, and/or composite features. Machine learning algorithms can be used to classify such objects, and image matching and verification can be adapted to the classification. Further, multiple modes of input can be integrated at various stages of the object recognition processing pipeline. These multi-modal inputs can include user feedback, additional images representing different perspectives of the object or specific regions of the object including a logo or text corresponding to the object, user behavior data, location, among others.

Object recognition of feature-sparse or texture-limited subject matter

Approaches provide for analyzing image data to determine and/or recognize text in the image data. The recognized text can be used to generate a search query that can be automatically submitted to a search engine without having to type the search query to identify a product (or related products) associated with the image. For example, a camera of a computing device can be used to capture a live camera view (or single images) an item. An application executing on the computing device (or at least in communication with the computing device) can analyze the image data of the live camera view to determine a set of keywords (e.g., identified text) based on visual features extracted from the image data. The keywords can be used to query an index of product titles, common search queries, among other indexed data to return a ranked list of search suggestions based on a relevance function. The relevance function can consider the ordering of the keywords to rank search suggestions more highly that contain the keywords having the same word order. Further, the relevance function can consider the confidence of the visual recognition of each keyword, the confidence of each search suggestion, customer impact, as well as other factors to determine the ranking of the search suggestions. The search suggestions can be further refined to ensure search results that the user will be more likely to view and/or purchase.

Visual search suggestions

Approaches include using a machine learning-based approach to generating search strings and refinements based on a specific item represented in an image. For example, a classifier that is trained on descriptions of images can be provided. An image that includes a representation of an item of interest is obtained. The image is analyzed using the classifier algorithm to determine a first term representing a visual characteristic of the image. Then, the image is analyzed again to determine a second term representing another visual characteristic of the image based at least in part on the first term. Additional terms can be determined to generate a description of the image, including characteristics of the item of interest. Based on the determined characteristics of the item of interest, a search query and one or more refinements can be generated.

Growing community of inventors

Sunnyvale, CA, United States of America

Edward Hsiao