System and method for procedurally synthesizing datasets of objects of interest for training machine-learning models

• US 9,996,771 B2
• Filed: 02/15/2016
• Issued: 06/12/2018
• Est. Priority Date: 02/15/2016
• Status: Active Grant

• Current Assignee: NVIDIA Corporation
• Original Assignee: NVIDIA Corporation
• Inventors: Clayton, Jesse, Glavtchev, Vladimir
• Primary Examiner(s): Seth, Manav
• Application Number:
  US15/043,697
• Publication Number:
  US 20170236013A1
• Time in Patent Office: 848 Days
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06F 18/254 of classification results, e.g. of results related to same input data
• CPC Class Codes: G06F 18/28 Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 5/046 Forward inferencing; Production systems
• CPC Class Codes: G06T 15/005 General purpose rendering architectures
• CPC Class Codes: G06V 10/772 Determining representative reference patterns, e.g. averaging or distorting patterns; Generating dictionaries
• CPC Class Codes: G06V 10/774 Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06V 20/58 Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
• CPC Class Codes: G06V 20/584 of vehicle lights or traffic lights
• CPC Class Codes: G06V 40/103 Static body considered as a whole, e.g. static pedestrian or occupant recognition
• CPC Class Codes: G06V 40/16 Human faces, e.g. facial parts, sketches or expressions
• CPC Class Codes: G06V 40/20 Movements or behaviour, e.g. gesture recognition recognition of facial expressions G06V40/16

Abstract

A system and method for procedurally synthesizing a training dataset for training a machine-learning model. In one embodiment, the system includes: (1) a training designer configured to describe variations in content of training images to be included in the training dataset and (2) an image definer coupled to the training designer, configured to generate training image definitions in accordance with the variations and transmit the training image definitions: to a 3D graphics engine for rendering into corresponding training images, and further to a ground truth generator for generating associated ground truth corresponding to the training images, the training images and the associated ground truth comprising the training dataset.

Claims

1. A method of procedurally synthesizing a training dataset for training a machine-learning model, comprising:

describing variations in content of training images to be included in said training dataset;

generating training image definitions in accordance with said variations in parallel;

employing said training image definitions to render corresponding training images in parallel;

further employing said training image definitions to generate associated ground truth in parallel; and

assembling said training images and associated ground truth into said training dataset.

2. The method as recited in claim 1 further comprising accepting input from a user establishing limits of said variations.

3. The method as recited in claim 1 further comprising refining said variations to eliminate variations that are physically impossible.

4. The method as recited in claim 1 wherein said training image definitions are expressed in a standard graphics language.

5. The method as recited in claim 1 wherein said employing comprises employing a conventional 3D graphics engine.

6. The method as recited in claim 1 wherein said further employing comprises further employing said training image definitions expressed in a graphics language.

7. The method as recited in claim 1 wherein said ground truth includes coordinates locating an object of interest in said training images.

8. A training dataset for training a machine-learning model procedurally synthesized by the process comprising:

describing variations in content of training images to be included in said training dataset;

generating training image definitions in accordance with said variations in parallel;

employing said training image definitions to render corresponding training images in parallel;

further employing said training image definitions to generate associated ground truth in parallel; and

assembling said training images and associated ground truth into said training dataset.

9. The training dataset as recited in claim 8 further comprising accepting input from a user establishing limits of said variations.

10. The training dataset as recited in claim 8 further comprising refining said variations to eliminate variations that are physically impossible.

11. The training dataset as recited in claim 8 wherein said employing comprises employing a conventional 3D graphics engine.

12. The training dataset as recited in claim 8 wherein said ground truth includes coordinates locating an object of interest in said training images.