Amazon Technologies, Inc.

A9.com, Inc.

Yifan Xing

Wei Xia

Yuanjun Xiong

Wei Li

Stefano Soatto

Paulo Ricardo Dos Santos Mendonca

Shuo Yang

Arnab Sanat Kumar Dhua

Meng Wang

Radek Grzeszczuk

Zheng Zhang

Himanshu Arora

Chun-Kai Wang

Siqi Deng

Paulo Mendonca

Tong He

Yongxin Wang

Tianjun Xiao

David Paul Wipf

Yantao Shen

Rahul Tewari

Techniques for performing visual clustering with a hierarchical graph neural network framework including a joint linkage prediction and density estimation graph model are described. Embodiments herein recurrently run the joint linkage prediction and density estimation graph model to generate intermediate clusters in multiple iterations (e.g., until convergence) to obtain a final clustering result. In certain embodiments, for each iteration, the input graph contains nodes that are merged from nodes assigned to intermediate clusters from the previous iteration. By using a small and fixed bandwidth k in each iteration, embodiments herein alleviate the sensitivity to the k selection for different clustering applications. Certain embodiments herein remove the tuning of a different k (e.g., k-bandwidth) for k-nearest neighbor graph construction over different clustering applications.

Hierarchical graph neural networks for visual clustering

Techniques for data synthesis for training datasets for machine learning applications are described. A first image of at least an object from a first viewpoint is obtained. The first image having associated first image metadata including a first location of a feature of the object in the first image. A model is generated from the first image, the model including a three-dimensional representation of the object. A second image is generated from the model, the second image including the object from a second viewpoint that is different from the first viewpoint. Second image metadata is generated, the second image metadata including a second location of the feature of the object in the second image, the second location corresponding to the first location adjusted for the difference between the second viewpoint and the first viewpoint.

Data synthesis using three-dimensional modeling

Techniques for building a backward compatible and backfill-free image search system are described. According to some embodiments, a backwards compatible training system trains a new embedding model to be backward compatible with the face embeddings (e.g., floating-point vectors) generated by a previous embedding model. In one embodiment, backwards compatible training uses a classifier of the previous embedding model as a form of constraint in the training of the new embedding model.

Backward compatible and backfill-free image search system

Approaches herein provide for multi-camera calibration and subsequent application in augmented reality applications. The approach obtains image data of a three-dimensional (3D) calibration object from different directions using the at least one camera. A determination is made for extrinsic parameters associated with a relative position of the 3D calibration object to the at least one camera. A further determination is made for intrinsic parameters associated with the at least one camera relative to the 3D calibration object facing in the different positions. Revised intrinsic parameters are generated from the intrinsic parameters. The revised intrinsic parameters provide a calibration output that may be applied to generate views of an item in an augmented reality application.

Camera calibration for augmented reality

Techniques for improving a 2D to 3D image reconstruction network machine learning model are described. In some instances, this includes performing at least two transformations of a 3D model to generate at least two rotated 3D models, the at least two transformations to rotate the 3D model about an axis away from a viewing direction of the single 2D image; rendering the at least two rotated 3D models as rendered 2D images; and retraining a 2D to 3D image reconstruction network machine learning model using corresponding pairs of rotated 3D models and rendered 2D images.

Self-supervised bootstrap for single image 3-D reconstruction

Objects can be rendered in three-dimensions and viewed and manipulated in an augmented reality environment. For example, contour information for an object represented in image data for a plurality of viewpoints can be obtained. A region of space that includes the object can be partitioned into a plurality of sub regions. A position of each sub region can be identified in the image data for each viewpoint. A value can be determined for each sub region indicating whether that sub region as projected in the image data is within the contour of the object. Once the total values for each sub region are determined, a visual hull technique can be used to generate a visual hull of the object based on the values. Thereafter, the visual hull can be used to generate a three-dimensional representation of the object for a number of different viewpoints.

Determining a visual hull of an object

Growing community of inventors

Seattle, WA, United States of America

Yifan Xing