Automated generation of training images

Patent metadata

Field	Value
Publication number	US-2021397893-A1
Application number	US-201917289649-A
Country	US
Kind code	A1
Filing date	Oct 30, 2019
Priority date	Oct 30, 2018
Publication date	Dec 23, 2021
Grant date	—

Abstract

A method for providing training images including receiving information about objects which is at least valid at a specific recording time. The method includes receiving an image of an area imaged at the specific recording time. The method includes estimating respective positions of the objects at the specific recording time based on the information and selecting estimated positions of respective objects from the estimated respective positions which fulfill a predefined selection criterion. The method includes generating training images by separating the imaged area into first and second pluralities of image tiles. Each of the first plurality of image tiles differs from each of the second plurality of image tiles. Each of the first plurality of image tiles images (depicts) a respective one or several of the selected positions. The method includes providing the training images. Further, a training image product and a device for providing training images are provided.

First claim

1 - 15 . (canceled)

16 . A method for providing training images, the method comprising:
receiving information about objects, wherein the information is at least valid at a specific recording time; receiving an image of an area, wherein the area is imaged at the specific recording time; estimating respective positions of the objects at the specific recording time based on the information; selecting estimated positions of respective objects from the estimated respective positions which fulfill a predefined selection criterion; generating training images by separating the imaged area into a first plurality of image tiles and a second plurality of image tiles, wherein each of the first plurality of image tiles differs from each of the second plurality of image tiles, and wherein each of the first plurality of image tiles images a respective one or several of the selected positions; and providing the training images.

17 . The method according to claim 16 , wherein the positions of the objects to be estimated are within the area at the specific recording time.

18 . The method according to claim 16 , wherein each of the first plurality of image tiles images one or several of the selected positions.

19 . The method according to claim 16 , wherein the predefined selection criterion is related to an error probability.

20 . The method according to claim 16 , wherein the image tiles of the second plurality of image tiles differ among each other.

21 . The method according to claim 16 , wherein the image tiles of the first plurality differ among each other.

22 . The method according to claim 16 , wherein the information comprises time dependent information about the objects.

23 . The method according to claim 16 , wherein the information comprises time independent information about the objects.

24 . The method according to claim 16 , wherein the information is transmitted by the objects.

25 . The method according to claim 16 , wherein the information comprises information about geographic coordinates, and wherein the method further comprises georeferencing the imaged area according to the information about the geographic coordinates.

26 . The method according to claim 16 , wherein the image tiles of the second plurality of image tiles image a part of the imaged area that does not contain an object of interest.

27 . A training image product as input to train machine learning for an image analysis system, the training image product comprising the training images provided by a method according to claim 16 .

28 . The training image product according to claim 27 ,
wherein the training images are separated into a first plurality of image tiles and a second plurality of image tiles, wherein each of the first plurality of image tiles comprises a respective state vector, and wherein components of the respective state vectors are associated with an object to be present in the respective image tile.

29 . The training image product according to claim 28 , wherein each of the second plurality of image tiles are labelled with the information that no object is present in the respective image tile.

30 . A device for providing training images, the device comprising:
a first receiving unit configured to receive information about objects, wherein the information is at least valid at a specific recording time; a second receiving unit configured to receive an image of an area, wherein the area is imaged at the specific recording time; a processing unit configured to estimate respective positions of the objects at the specific recording time based on the information, wherein the processing unit is further configured to select estimated positions of respective objects from the estimated respective positions which fulfill a predefined selection criterion, and wherein the processing unit is further configured to generate training images by separating the imaged area into a first plurality of image tiles and a second plurality of image tiles, wherein each of the first plurality of image tiles differs from each of the second plurality of image tiles, and wherein each of the first plurality of image tiles images a respective one or several of the selected positions; and a transmitting unit configured to provide the training images.

Assignees

• Airbus Defence & Space Gmbh

Inventors

• Mohrdieck Camilla
• Wiedmann Michael
• Zimmermann Reiner

Classifications

• G06V10/774 Primary

  Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06V20/13

  Satellite images · CPC title

• G06F18/2148

  characterised by the process organisation or structure, e.g. boosting cascade · CPC title

• G06F18/2115

  by evaluating different subsets according to an optimisation criterion, e.g. class separability, forward selection or backward elimination · CPC title

• G06T7/11

  Region-based segmentation · CPC title