Amazon Technologies, Inc.

Other

Francesco Callari

Jean-Guillaume Dominique Durand

Pradeep Krishna Yarlagadda

Ishay Kamon

Pradeep Krishna Krishna Yarlagadda

Tatiana Glozman

Harshul Gupta

Systems and methods to automatically generate training data including airborne object trajectory data may receive real trajectory data from within a first environment, receive imaging data associated with a second environment, and transform and/or modify the real trajectory data to synthetic trajectory data for the second environment. Then, the synthetic trajectory data may be superimposed within the imaging data of the second environment. In addition, images of an airborne object may be rendered along the synthetic trajectory data to generate training data that may be used to train machine learning models or algorithms for various computer vision applications.

Automated generation of training data including airborne object trajectory data

Systems and methods to automatically generate training data for machine learning models may include an imaging device to capture imaging data, an image processing or rendering system to receive the imaging data and render a three-dimensional model of an object of interest overlaying the imaging data, an automatic mask extraction or generation system to extract or determine a mask, label, or annotation associated with the three-dimensional model and a plurality of pixels associated with the object of interest from a perspective of the imaging device, and a machine learning model to receive the imaging data and the mask as training data.

Automatic generation of training data for instance segmentation algorithms

Data for training a machine learning algorithm to detect airborne objects is generated by capturing background images using a camera aboard an aerial vehicle, and generating a trajectory of a synthetic object, such another aerial vehicle, in three-dimensional space. The trajectory is projected into an image plane of the camera, determined from a pose of the camera calculated using inertial measurement unit data captured by the aerial vehicle. Images of the synthetic object may be rendered based on locations of the trajectory in the image plane at specific times. Pixel-accurate locations for the rendered images may be determined by calculating a homography from consecutive images captured using the camera, and adjusting locations of the trajectory using the homography. The rendered images may be blended into the images captured by the camera at such locations, and used to train a machine learning algorithm.

Generating data for training vision-based algorithms to detect airborne objects

Described are systems and methods for generating training data that is used to train a machine learning system to detect moving objects represented in sensor data. The system and methods utilize position data received from a target vehicle to determine data points within sensor data that represents that target vehicle. For example, a station at a known location may receive Automatic Dependent Surveillance-Broadcast ('ADS-B') data (position data) corresponding to a target vehicle that is within the field of view of a station sensor, such as a camera. The position data may then be correlated with the sensor data and projected into the sensor data to determine data points within the sensor data that represent the target vehicle. Those data points are then labeled to indicate the location, size, and/or shape of the target vehicle as represented in the sensor data, thereby producing training that may be provided to train a machine learning algorithm or system to detect moving objects, such as aircraft.

Producing training data for machine learning

Systems and methods are provided herein for managing a set of autonomous vehicles (AVs) configured to perform delivery tasks and computing tasks. Computing tasks can be performed such as training a model and/or calculating an incremental update for the model. As additional training data is obtained, a subset of AVs may be managed as a distributed computing cluster and assigned a computing task such as training or calculating an incremental update for the model or any suitable computing task. Corresponding data computed by the subset of AVs of the cluster (e.g., the retrained model, updated model parameters corresponding to the updated model, etc.) may be received and stored or transmitted (e.g., the computing task requestor, to the AVs, etc.) for subsequent use (e.g., for subsequent delivery tasks).

Techniques for managing processing resources

Embodiments of the invention relate to a customizable modular pet house and entertainment system. In one general embodiment of the invention, a pet house comprises a module formed from a sheet of self-supporting material. The module comprises a top panel, a bottom panel, and four side panels. At least one of the panels may have at least one of a removable transition panel, a removable door panel, and perimeter of a removable window panel. Optional accessories include a top accessory panel, a side accessory panel, and an accessory bracket. More than one modular pet house may be joined together to form an extended structure. Embodiments of the invention have interior spaces and apertures sized for admittance of a common house cat.

Growing community of inventors

Seattle, WA, United States of America

Francesco Callari