Training and deploying pose regressions in neural networks in autonomous machines

Patent metadata

Field	Value
Publication number	US-12536693-B2
Application number	US-202318465683-A
Country	US
Kind code	B2
Filing date	Sep 12, 2023
Priority date	Sep 9, 2016
Publication date	Jan 27, 2026
Grant date	Jan 27, 2026

Abstract

A mechanism is described for facilitating training and deploying of pose regression in neural networks in autonomous machines. A method, as described herein, includes facilitating capturing, by an image capturing device of a computing device, one or more images of one or more objects, where the one or more images include one or more training images associated with a neural network. The method may further include continuously estimating, in real-time, a present orientation of the computing device, where estimating includes continuously detecting a real-time view field as viewed by the image capturing device and based on the one or more images. The method may further include applying pose regression relating to the image capturing device using the real-time view field.

First claim

What is claimed is:

1 . An apparatus comprising:
processing circuitry coupled to a memory, the processing circuitry to: form rotation matrix or rotation quaternion corresponding to rotation representations of one or more images as captured by an image capturing device; transition the rotation matrix or the rotation quaternion to one or more decomposed angle representations based on one or more angle estimations; and apply pose regression relating to the image capturing device based on a real-time view field, wherein the one or more decomposed angle representations comprise one or more angles corresponding to one or more movements of the image capturing device, wherein the pose regression is associated with a prediction error such that the prediction error is estimated for and applied to the pose regression.

2 . The apparatus of claim 1 , wherein the one or more angles are presented as one or more of cos(yaw), sin(yaw), cos(pitch), sin(pitch), cos(roll), and sin(roll), wherein the image capturing device to capture the one or more images of one or more objects, wherein the one or more images include one or more training images associated with a neural network.

3 . The apparatus of claim 1 , wherein the processing circuitry is further to:
continuously estimate, in real-time, a present orientation of the apparatus, wherein to continuously estimate includes to continuously detect the real-time view field as viewed by the image capturing device and based on the one or more images.

4 . The apparatus of claim 3 , wherein the view field to provide at least one of translations representing global coordinates and rotations representing movements of the image capturing device along its axes, wherein to apply pose regression includes to adjust the present orientation of the apparatus to facilitate accurate capturing of input data and offering of output results associated with workings of the neural network.

5 . The apparatus of claim 1 , wherein the prediction error is estimated, in real-time, based on a difference between two consecutive rotations, wherein the difference is regarded as the prediction error, wherein, once the prediction error is applied to the pose regression, the pose regression is adjusted in accordance with the prediction error.

6 . The apparatus of claim 2 , wherein the input capturing device comprises at least one of one or more cameras, one or more robot eyes, one or more microphones, and one or more sensors, wherein the apparatus comprises an autonomous machine or an artificially intelligent agent, wherein the autonomous machine includes at least one of one or more robots, one or more self-driving vehicles, and one or more self-operating equipment, wherein the processing circuitry comprises one or more of graphics processing circuitry or application processing circuitry.

7 . A method comprising
forming, by a processing circuitry of a computing device, rotation matrix or rotation quaternion corresponding to rotation representations of one or more images as captured by an image capturing device; transitioning, by the processing circuitry, the rotation matrix or the rotation quaternion to one or more decomposed angle representations based on one or more angle estimations; and applying pose regression relating to the image capturing device based on a real-time view field, wherein the one or more decomposed angle representations comprise one or more angles corresponding to one or more movements of the image capturing device, wherein the pose regression is associated with a prediction error such that the prediction error is estimated for and applied to the pose regression.

8 . The method of claim 7 , wherein the one or more angles are presented as one or more of cos(yaw), sin(yaw), cos(pitch), sin(pitch), cos(roll), and sin(roll), wherein the image capturing device to capture the one or more images of one or more objects, wherein the one or more images include one or more training images associated with a neural network.

9 . The method of claim 7 , further comprising:
continuously estimating, in real-time, a present orientation of the apparatus, wherein continuously estimating includes continuously detecting the real-time view field as viewed by the image capturing device and based on the one or more images.

10 . The method of claim 9 , wherein the view field to provide at least one of translations representing global coordinates and rotations representing movements of the image capturing device along its axes, wherein to apply pose regression includes to adjust the present orientation of the apparatus to facilitate accurate capturing of input data and offering of output results associated with workings of the neural network.

11 . The method of claim 7 , wherein the prediction error is estimated, in real-time, based on a difference between two consecutive rotations, wherein, once the prediction error is applied to the pose regression, the pose regression is adjusted in accordance with the prediction error.

12 . The method of claim 8 , wherein the input capturing device comprises at least one of one or more cameras, one or more robot eyes, one or more microphones, and one or more sensors, wherein the apparatus comprises an autonomous machine or an artificially intelligent agent, wherein the autonomous machine includes at least one of one or more robots, one or more self-driving vehicles, and one or more self-operating equipment, wherein the processing circuitry comprises one or more of graphics processing circuitry or application processing circuitry.

13 . At least one non-transitory computer-readable medium having stored thereon instructions which, when executed, cause a computing device to perform operations comprising:
forming rotation matrix or rotation quaternion corresponding to rotation representations of one or more images as captured by an image capturing device; transitioning the rotation matrix or the rotation quaternion to one or more decomposed angle representations based on one or more angle estimations; and applying pose regression relating to the image capturing device based on a real-time view field, wherein the one or more decomposed angle representations comprise one or more angles corresponding to one or more movements of the image capturing device, wherein the pose regression is associated with a prediction error such that the prediction error is estimated for and applied to the pose regression.

14 . The non-transitory computer-readable medium of claim 13 , wherein the one or more angles are presented as one or more of cos(yaw), sin(yaw), cos(pitch), sin(pitch), cos(roll), and sin(roll), wherein the image capturing device to capture the one or more images of one or more objects, wherein the one or more images include one or more training images associated with a neural network.

15 . The non-transitory computer-readable medium of claim 13 , wherein the operations further comprise:
continuously estimating, in real-time, a present orientation of the apparatus, wherein continuously estimating includes continuously detecting the real-time view field as viewed by the image capturing device and based on the one or more images.

16 . The non-transitory computer-readable medium of claim 14 , wherein the view field to provide at least one of translations representing global coordinates and rotations representing movements of the image capturing device along its axes, wherein to apply pose regression includes to adjust the present orientation of the apparatus to facilitate accurate capturing of input data and offering of output results associated with

Assignees

• Intel Corp

Inventors

• Ma Liwei

Classifications

• G06T2207/30244

  Camera pose · CPC title

• G06T2207/20084

  Artificial neural networks [ANN] · CPC title

• G06T2207/20081

  Training; Learning · CPC title

• G06N3/08 Primary

  Learning methods · CPC title

• G06T7/73 Primary

  using feature-based methods · CPC title