Method and system for generating training data

Patent metadata

Field	Value
Publication number	US-12067080-B2
Application number	US-202117375356-A
Country	US
Kind code	B2
Filing date	Jul 14, 2021
Priority date	Jul 14, 2020
Publication date	Aug 20, 2024
Grant date	Aug 20, 2024

Abstract

A method for generating training data can include: determining a set of images; determining a set of masks based on the images; determining a first mesh based on the set of masks; optionally determining a refined mesh by recomputing the first mesh; optionally determining one or more faces of the refined mesh; optionally adding one or more keypoints to the refined mesh; optionally determining a material property set for the object; optionally generating a full object mesh; determining one or more scenes; optionally determining training data based on the one or more scenes; optionally training one or more object detectors using the training data; and detecting one or more objects using the trained object detector.

First claim

We claim:

1. A method, comprising:
capturing a set of images that depict an object;
segmenting an image of the set of images to determine a volumetric object segment based on back lit images of the object from the set of images;
determining an object mesh based on the volumetric object segment;
determining multiple object faces of the object mesh; and
adding multiple keypoints to each object face of the object mesh, wherein each keypoint is associated with an object face identifier for the respective object face.

2. The method of claim 1 , wherein the set of images comprises depth images that depict a silhouette of the object, wherein the depth images comprise surface normals for the keypoints in each image, and the method further comprises determining refined surface normals using a bidirectional reflectance distribution function.

3. The method of claim 2 , further comprising rendering a synthetic image using the object mesh and the refined surface normals.

4. The method of claim 2 , wherein the set of images comprises a front lit image captured at a same position as the back lit image, and wherein rendering the synthetic image comprises texturing the object mesh using the front lit image.

5. The method of claim 1 , further comprising:
generating multiple refined meshes for different sides of the object; and
generating a full mesh by aligning the multiple refined meshes and combining the aligned multiple refined meshes.

6. The method of claim 1 , wherein each object face of the multiple object faces is determined based on surface normal continuity of the object mesh, and an object component identifier is assigned to keypoints that lie on the object face.

7. The method of claim 6 , wherein the keypoints are placed on each object face to maximize keypoint coverage of the object face.

8. The method of claim 1 , wherein the set of images are captured using an image capture environment, wherein the image capture environment comprises a housing that defines a measurement volume, and a first set of light emitting elements configured to front light the measurement volume, and a second set of light emitting elements configured to back light the measurement volume.

9. The method of claim 8 , wherein the first set of light emitting elements and the second set of light emitting elements are distributed about an interior of the housing.

10. A system, comprising:
one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising:
capturing a set of images depicting an object;
determining a set of masks based on the set of images;
determining an object mesh based on the set of masks;
determining multiple object faces of the object mesh; and
adding a plurality of keypoints to each object face of the object mesh, wherein each keypoint is associated with an object face identifier for the respective object face.

11. The system of claim 10 , further comprising an image capture environment comprising:
a housing defining an enclosed measurement volume;
a pedestal mounted to the housing within the measurement volume and configured to support an object in the measurement volume;
a set of calibration features mounted to the pedestal;
an imaging system configured to capture the set of images;
an imaging system actuator mounting the imaging system and configured to move the imaging system along a predefined trajectory relative to the pedestal; and
a plurality of light emitting elements mounted to the housing, wherein the light emitting elements are distributed about the interior of the housing.

12. The system of claim 11 , wherein the plurality of light emitting elements comprises a front and back subset of light emitting elements configured to front-light and back-light the measurement volume, wherein the images captured when the measurement volume is back lit are used to generate the object mesh, and wherein the images that are captured when the measurement volume is front lit are used to render a synthetic image.

13. The system of claim 11 , wherein the pedestal rotates during image capture.

14. The system of claim 11 , wherein the calibration pattern is mounted to a side of the pedestal.

15. The system of claim 11 , wherein the light emitting elements are evenly distributed about the measurement volume.

16. The system of claim 11 , wherein the imaging system actuator is configured to actuate the camera to capture a plurality of views of the object, wherein the plurality of views cooperatively form a photo dome of the object.

17. The system of claim 10 , wherein the operations further comprise rendering a synthetic image using the object mesh, wherein the synthetic image is rendered using a physics module configured to simulate a scene using multiple instances of the object mesh.

18. The system of claim 17 , wherein the synthetic image is associated with unoccluded keypoints from the plurality that are unoccluded in the synthetic image.

19. The system of claim 17 , wherein the synthetic image is used as input to train an object detector to identify unoccluded object keypoints in the synthetic image.

20. The system of claim 10 , wherein the set of images comprises a back lit image, and wherein the set of masks is determined based on the back lit image.

21. The method of claim 1 , further comprising training an object detector using synthetic images and the generated keypoints as training data.

22. The system of claim 10 , wherein the operations further comprise training an object detector using synthetic images and the generated keypoints as training data.

Assignees

• Intrinsic Innovation Llc

Inventors

• Eyjolfsdottir Eyrun
• Stark Michael
• Carra Marco
• Gangwar Manushree
• Schlegel Alex
• Yuvaraj Karthikeyan
• Bandy Scott
• So Peter
• Whitaker Lucas
• Swaminathan Sivaramakrishnan

Classifications

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• H04N23/695

  Control of camera direction for changing a field of view, e.g. pan, tilt or based on tracking of objects · CPC title

• H04N23/56

  provided with illuminating means · CPC title

• G06T17/205

  Re-meshing · CPC title