Techniques for generating subjective style comparison metrics for B-reps of 3D CAD objects

Patent metadata

Field	Value
Publication number	US-12380256-B2
Application number	US-202117523725-A
Country	US
Kind code	B2
Filing date	Nov 10, 2021
Priority date	Nov 13, 2020
Publication date	Aug 5, 2025
Grant date	Aug 5, 2025

Abstract

In various embodiments, a style comparison metric application generates a style comparison metric for pairs of different three dimensional (3D) computer-aided design (CAD) objects. In operation, the style comparison metric application executes a trained neural network any number of times to map 3D CAD objects to feature maps. Based on the feature maps, the style comparison metric application computes style signals. The style comparison metric application determines values for weights based on the style signals. The style comparison metric application generates the style comparison metric based on the weights and a parameterized style comparison metric.

First claim

What is claimed is:

1. A computer-implemented method for generating a style comparison metric for pairs of different three dimensional (3D) computer-aided design (CAD) objects, the method comprising:
executing a trained neural network one or more times to map inputs of the trained neural network comprising a plurality of 3D CAD objects to outputs of the trained neural network comprising a plurality of feature maps, wherein the trained neural network is generated using unsupervised learning techniques that do not receive labeled training data as input;
computing a plurality of style signals based on the plurality of feature maps;
determining a plurality of values for a plurality of weights based on the plurality of style signals, wherein a parameterized style comparison metric combines a plurality of style distances based on the plurality of weights; and
generating the style comparison metric based on the plurality of weights and the parameterized style comparison metric.

2. The computer-implemented method of claim 1 , wherein a first 3D CAD object included in the plurality of 3D CAD objects comprises a positive example of a target style.

3. The computer-implemented method of claim 2 , wherein a second 3D CAD object included in the plurality of 3D CAD objects comprises a negative example of the target style.

4. The computer-implemented method of claim 1 , further comprising:
randomly selecting a first 3D CAD object from a data set of 3D CAD objects and including the first 3D CAD object in the plurality of 3D CAD objects; and
designating the first 3D CAD object as a negative example of a target style.

5. The computer-implemented method of claim 1 , wherein determining the plurality of values for the plurality of weights comprises:
generating a parameterized loss based on the plurality of style signals and the plurality of weights; and
performing one or more optimization operations on the parameterized loss to determine the plurality of values for the plurality of weights.

6. The computer-implemented method of claim 1 , wherein computing the plurality of style signals comprises:
performing one or more per-face recentering operations on a first feature map included in the plurality of feature maps to generate a first normalized feature map; and
extracting at least a portion of a Gram matrix of the first normalized feature map to generate a first style signal included in the plurality of style signals.

7. The computer-implemented method of claim 1 , wherein generating the style comparison metric comprises replacing a first weight that is specified in the parameterized style comparison metric and included in the plurality of weights with a first value included in the plurality of values.

8. The computer-implemented method of claim 1 , wherein the trained neural network comprises a trained encoder neural network that is configured to map boundary-representations (B-reps) of the plurality of 3D CAD objects to the plurality of feature maps, and further comprising:
generating a first representation based on a first B-rep that represents a first 3D CAD object included in the plurality of 3D CAD objects; and
inputting the first representation into the trained encoder neural network.

9. The computer-implemented method of claim 1 , wherein a representation of a first 3D CAD object included in the plurality of 3D CAD objects comprises a boundary-representation (B-rep), a 3D mesh, or a 3D point cloud.

10. The computer-implemented method of claim 1 , wherein the trained neural network comprises a trained encoder or a trained classifier.

11. One or more non-transitory computer readable media including instructions that, when executed by one or more processors, cause the one or more processors to generate a style comparison metric for pairs of different three dimensional (3D) computer-aided design (CAD) objects by performing the steps of:
executing a trained neural network one or more times to map inputs of the trained neural network comprising a plurality of 3D CAD objects to outputs of the trained neural network comprising a plurality of feature maps, wherein the trained neural network is generated using unsupervised learning techniques that do not receive labeled training data as input;
computing a plurality of style signals based on the plurality of feature maps;
determining a plurality of values for a plurality of weights based on the plurality of style signals, wherein a parameterized style comparison metric combines a plurality of style distances based on the plurality of weights; and
generating the style comparison metric based on the plurality of weights and the parameterized style comparison metric.

12. The one or more non-transitory computer readable media of claim 11 , wherein a first 3D CAD object included in the plurality of 3D CAD objects comprises a user-specified positive example of a target style.

13. The one or more non-transitory computer readable media of claim 12 , wherein a second 3D CAD object included in the plurality of 3D CAD objects comprises a negative example of the target style.

14. The one or more non-transitory computer readable media of claim 11 , further comprising:
randomly selecting a first 3D CAD object from a data set of 3D CAD objects and including the first 3D CAD object in the plurality of 3D CAD objects; and
designating the first 3D CAD object as a negative example of a target style.

15. The one or more non-transitory computer readable media of claim 11 , wherein determining the plurality of values for the plurality of weights comprises:
generating a parameterized loss based on the plurality of style signals and the plurality of weights; and
performing one or more optimization operations on the parameterized loss to determine the plurality of values for the plurality of weights.

16. The one or more non-transitory computer readable media of claim 11 , wherein computing the plurality of style signals comprises:
performing one or more instance normalization operations on a first feature map included in the plurality of feature maps to generate a first normalized feature map; and
extracting at least a portion of a Gram matrix of the first normalized feature map to generate a first style signal included in the plurality of style signals.

17. The one or more non-transitory computer readable media of claim 11 , wherein generating the style comparison metric comprises replacing a first weight that is specified in the parameterized style comparison metric and included in the plurality of weights with a first value included in the plurality of values.

18. The one or more non-transitory computer readable media of claim 11 , wherein the trained neural network includes a trained face encoder and a trained graph encoder.

19. The one or more non-transitory computer readable media of claim 11 , wherein a representation of a first 3D CAD object included in the plurality of 3D CAD objects comprises a boundary-representation (B-rep), a 3D mesh, or a 3D point cloud.

20. A system comprising:
one or more memories storing instructions; and
one or more processors coupled to the one or more memories that, when
executing the instructions, perform the steps of:
executing a trained neural network one or more times to map inputs of the trained neural network comprising a plurality of 3D CAD objects to outputs of the trained neural network comprising a plurality of feature maps, wherein the trained neural network is generated using unsupervised learning techniques that do not receive labeled t

Assignees

• Autodesk Inc

Inventors

• Meltzer Peter
• Khas Ahmadi Amir Hosein
• Jayaraman Pradeep Kumar
• Lambourne Joseph George
• Sanghi Aditya
• Shayani Hooman

Classifications

• G06N3/0895

  Weakly supervised learning, e.g. semi-supervised or self-supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/088

  Non-supervised learning, e.g. competitive learning · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06F30/27 Primary

  using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model · CPC title