Amazon Technologies, Inc.

Leonid Pishchulin

Jean Laurent Guigues

Gerard Guy Medioni

Prithviraj Banerjee

Jayakrishnan Kumar Eledath

Dinesh Reddy Narapureddy

Ning Xu

Maxim Berman

Disclosed are systems and methods to perform neural architecture search ('NAS') that automatically optimizes for the number of channels to allocate to each layer of a deep neural network. Some implementations include a pairwise slimming that includes a global optimization step. Likewise, in some implementations, a bias toward a region of interest may be applied to channel path selection during training.

Slimmable neural network architecture search optimization

Actors may be detected and tracked within a scene using multiple imaging devices provided in a network that are aligned with fields of view that overlap at least in part. Processors operating on the imaging devices may evaluate the images using one or more classifiers to recognize body parts within the images, and to associate the body parts with a common actor within the scene. Each of the imaging devices may generate records of the positions of the body parts and provide such records to a central server, that may correlate body parts appearing within images captured by two or more of the imaging devices and generate a three-dimensional model of an actor based on positions of the body parts. Motion of the body parts may be tracked in subsequent images, and the model of the actor may be updated based on the motion.

Generating tracklets from digital imagery

Poses or gestures of actors within a scene may be detected and tracked using multiple imaging devices aligned with fields of view that overlap at least in part. Images captured by the imaging devices may be synchronized and provided to a classifier to recognize body parts within the images, and score maps indicative of locations of peak probabilities that the images include the respective body parts may be generated. Locations of peak values within the score maps may be correlated with one another to confirm that a given body part is depicted in two or more fields of view, and vectors indicative of distances to or ranges of motion of body parts, with respect to the given body part, may be generated. Motion of the body parts may be tracked in subsequent images, and a virtual model of the body parts may be generated and updated based on the motion.

Recognizing and tracking poses using digital imagery captured from multiple fields of view

Devices and techniques are generally described for articulated three-dimensional pose tracking. In some examples, a plurality of frames of image data captured by one or more cameras may be received. First feature data representing the plurality of frames of image data may be determined using a backbone network. The first feature data may be projected into three-dimensional (3D) space. In some examples, 3D location data describing respective 3D locations of one or more persons represented by the first feature data projected in the 3D space may be determined. The first feature data and the 3D location data may be sent to a four-dimensional (4D) convolutional neural network (CNN). The 4D CNN may generate second feature data comprising respective 3D representations of the one or more persons. Three dimensional pose data representing articulated 3D pose information for the one or more persons may be generated.

Growing community of inventors

Seattle, WA, United States of America

Leonid Pishchulin