US 11,961,281 B1Granted

Training and using computer vision model for item segmentations in images

AssigneeAmazon Technologies, Inc.

InventorsTaewan Kim, Jesse Norman Clark, Onkar Jayant Dabeer

Life of this patent

Estimating the term…

Loading patent…

Abstract

Techniques for training a machine-learning model are described. In an example, a computer generates a first pseudo-label indicating a first mask associated with a first object detected by a first machine-learning model in a first training image. A transformed image of the first training image can be generated using a transformation. Based on the transformation, a second pseudo-label indicating a second mask detected in the transformed image and corresponding to the first mask can be determined. A second machine-learning model can be trained using the second pseudo-label. The trained, second machine-learning model can detect a third mask associated with a second object based on a second image.

Claims20

Dependent claims

One or more computer-readable storage media storing instructions that, upon execution on a system, configure the system to perform operations comprising:

training, by at least using labeled data, a first machine learning model, the labeled data indicating a first mask associated with a first object present in a first training image, the first object associated with an item package classification;
generating, by at least using a second training image as input to the first machine learning model, a first pseudo-label indicating a second mask associated with a second object detected by the first machine learning model in the second training image, the second object associated with the item package classification;
generating, by at least using a transformation, a transformed image of the second training image;
determining, based at least in part on the transformation, a second pseudo-label indicating a third mask that is detected in the transformed image and that corresponds to the second mask;
training, by at least using the labeled data and the second pseudo-label, a second machine learning model;
receiving a third image that shows an item package;
determining, by at least using the third image as input to the second machine learning model, a fourth mask associated with a third object detected in the third image, the third object corresponding to the item package;
and
causing a manipulation by a robotic manipulator of the item package based at least in part on the fourth mask.

6 more elements

A computer-implemented method, comprising:

generating, by at least using a first training image as input to a first machine learning model, a first pseudo-label indicating a first mask associated with a first object detected by the first machine learning model in the first training image;
generating, by at least using a transformation, a transformed image of the first training image;
determining, based at least in part on the transformation, a second pseudo-label indicating a second mask that is detected in the transformed image and that corresponds to the first mask;
and
training, by at least using the second pseudo-label, a second machine learning model, wherein a third mask associated with a second object is detected by the second machine learning model based at least in part on a second image, wherein the second image is input to the second machine learning model and comprises the second object.

2 more elements

A system comprising:

one or more processors;
and
one or more memory storing computer-readable instructions that, upon execution by the one or more processors, configure the system to:
generate, by at least using a first training image as input to a first machine learning model, a first pseudo-label indicating a first mask associated with a first object detected by the first machine learning model in the first training image;
generate, by at least using a transformation, a transformed image of the first training image;
determine, based at least in part on the transformation, a second pseudo-label indicating a second mask that is detected in the transformed image and that corresponds to the first mask;
and
train, by at least using the second pseudo-label, a second machine learning model, wherein a third mask associated with a second object is detected by the second machine learning model based at least in part on a second image, wherein the second image is input to the second machine learning model and comprises the second object.

5 more elements

Dependencies are inferred from the claim text.

Facts

Application: 17/545,119Filed 8 Dec 2021
Granted: 16 Apr 2024
Term ends: Not estimatedThe source records don’t support an estimate.

Classification13 codes

G05B 19/4182
Characterised by the cooperation between machine tools, manipulators and conveyor or other workpiece supply system, workcell: manipulators and…
G06V 10/774
Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or…:…
B25J 9/163
Characterised by the control loop: learning, adaptive, model based, rule based expert control
B25J 9/1697
Characterised by use of sensors other than normal servo-feedback from position, speed or acceleration sensors, perception control, multi-sensor…:…