Apple Inc.

Joshua M Susskind

Shih-Yu Sun

Chen Huang

Feng Tang

Nitish Srivastava

Shuangfei Zhai

Miguel Angel Bautista Martin

Walter A Talbott

Russell Y Webb

Qi Shan

Tomas Jon Pfister

Micah P Kalscheur

Cuneyt Oncel Tuzel

Seyed Hesameddin Najafi Shoushtari

Aditya Sankar

Hanlin Goh

Ashish Shrivastava

Terrance Devries

Frankie Lu

Walter A Talbot

Robert Alex Colburn

Emilien Dupont

Attention-free transformers are disclosed. Various implementations of attention-free transformers include a gating and pooling operation that allows the attention-free transformers to provide comparable or better results to those of a standard attention-based transformer, with improved efficiency and reduced computational complexity with respect to space and time.

Attention free transformer

Implementations of the subject technology relate to generative scene networks (GSNs) that are able to generate realistic scenes that can be rendered from a free moving camera at any location and orientation. A GSN may be implemented using a global generator and a locally conditioned radiance field. GSNs may employ a spatial latent representation as conditioning for a grid of locally conditioned radiance fields, and may be trained using an adversarial learning framework. Inverting a GSN may allow free navigation of a generated scene conditioned on one or more observations.

Generative scene networks

Feature descriptor matching is reformulated into a graph-matching problem. Keypoints from a query image and a reference image are initially matched and filtered based on the match. For a given keypoint, a feature graph is constructed based on neighboring keypoints surrounding the given keypoint. The feature graph is compared to a corresponding feature graph of a reference image for the matched keypoint. Relocalization data is obtained based on the comparison.

Contextual matching

The subject technology provides a framework for learning neural scene representations directly from images, without three-dimensional (3D) supervision, by a machine-learning model. In the disclosed systems and methods, 3D structure can be imposed by ensuring that the learned representation transforms like a real 3D scene. For example, a loss function can be provided which enforces equivariance of the scene representation with respect to 3D rotations. Because naive tensor rotations may not be used to define models that are equivariant with respect to 3D rotations, a new operation called an invertible shear rotation is disclosed, which has the desired equivariance property. In some implementations, the model can be used to generate a 3D representation, such as mesh, of an object from an image of the object.

Neural rendering

Techniques for coding sets of images with neural networks include transforming a first image of a set of images into coefficients with an encoder neural network, encoding a group of the coefficients as an integer patch index into coding table of table entries each having vectors of coefficients, and storing a collection of patch indices as a first coded image. The encoder neural network may be configured with encoder weights determined by jointly with corresponding decoder weights of a decoder neural network on the set of images.

Neural network based image set compression

A generative network may be learned in an adversarial setting with a goal of modifying synthetic data such that a discriminative network may not be able to reliably tell the difference between refined synthetic data and real data. The generative network and discriminative network may work together to learn how to produce more realistic synthetic data with reduced computational cost. The generative network may iteratively learn a function that synthetic data with a goal of generating refined synthetic data that is more difficult for the discriminative network to differentiate from real data, while the discriminative network may be configured to iteratively learn a function that classifies data as either synthetic or real. Over multiple iterations, the generative network may learn to refine the synthetic data to produce refined synthetic data on which other machine learning models may be trained.

Generating more realistic synthetic data with adversarial nets

Systems and processes for metric learning distillation are disclosed herein. In accordance with one example, a method includes, at an electronic device, at an electronic device having one or more processors and memory, receiving a first plurality of vectors from a first model, receiving a second plurality of vectors from a second model, determining a first plurality of vector distances based on the first plurality of vectors, generating a first matrix based on the first plurality of vector distances, determining a second plurality of vector distances based on the second plurality of vectors, generating a second matrix based on the second plurality of vector distances, comparing the first matrix with the second matrix, and adjusting the second model based on the comparison of the first matrix and the second matrix.

Model and ensemble compression for metric learning

A facial recognition authentication process operating on a device may capture an image of a user using a camera on the device. The facial recognition authentication process may include operating a full face facial recognition authentication process on the captured image or operating a partial face facial recognition authentication process on the captured image. The process may determine which process to operate (either full face or partial face) based on an assessment of an amount of occlusion in the captured image. The partial face facial recognition authentication process may be operated when there is at least some occlusion of selected features (e.g., nose and/or mouth) on the user's face in the captured image.

Periocular facial recognition switching

A facial recognition authentication process may utilize images of a user's face that are captured while the user is being illuminated using both flood infrared illumination and patterned illumination (e.g., speckle pattern illumination). As the user's face is illuminated by both flood infrared illumination and patterned illumination, the captured images may include both flood infrared illumination data and depth map image data. Flood infrared illumination data may be generated from the images to assess two-dimensional features of the user in the captured images. Depth map image data may be generated from the pattern data in the images to assess three-dimensional (depth) features of the user in the captured images. The flood infrared illumination data and the depth map image data may be used separately by facial recognition authentication process to attempt to authenticate the user in the captured images as an authorized user of the device.

Growing community of inventors

San Jose, CA, United States of America

Joshua M Susskind