Generation of synthetic images for training a neural network model
Training deep neural networks requires a large amount of labeled training data. Conventionally, labeled training data is generated by gathering real images that are manually labelled which is very time-consuming. Instead of manually labelling a training dataset, domain randomization technique is used generate training data that is automatically labeled. The generated training data may be used to train neural networks for object detection and segmentation (labelling) tasks. In an embodiment, the generated training data includes synthetic input images generated by rendering three-dimensional (3D) objects of interest in a 3D scene. In an embodiment, the generated training data includes synthetic input images generated by rendering 3D objects of interest on a 2D background image. The 3D objects of interest are objects that a neural network is trained to detect and/or label.
1. A computer-implemented method, comprising:
rendering a three-dimensional (3D) object of interest to produce a rendered image of the object of interest, wherein an input image comprises the rendered image of the object of interest and a background image;
rendering one or more 3D geometric shapes to produce one or more rendered geometric shapes, wherein the one or more rendered geometric shapes are included in the input image;
computing task-specific training data corresponding to the object of interest; and
including the task-specific training data corresponding to the object of interest and the input image as a test pair in a training dataset for training a neural network to detect the object of interest and ignore the one or more rendered geometric shapes.
2. The computer-implemented method of claim 1 , further comprising receiving a position and orientation of a light for rendering the 3D object of interest.
3. The computer-implemented method of claim 2 , further comprising receiving a color of the light for rendering the 3D object of interest.
4. The computer-implemented method of claim 2 , further comprising receiving an intensity of the light for rendering the 3D object of interest.
5. The computer-implemented method of claim 1 , wherein rendering the one or more 3D geometric shapes comprises applying a texture map to at least one of the one or more 3D geometric shapes.
6. The computer-implemented method of claim 1 , further comprising: rendering an additional 3D object of interest to produce an additional rendered object of interest, wherein the rendered additional object of interest is included in the input image; and computing additional task-specific training data corresponding to the additional object of interest, wherein the additional task-specific training data is included in the task pair.
7. The computer-implemented method of claim 6 , wherein the additional rendered object of interest occludes a portion of the rendered object of interest, and when the portion is greater than a predetermined threshold value, modifying the task-specific training data corresponding to the object of interest.
8. The computer-implemented method of claim 1 , wherein rendering the 3D object of interest comprises rendering the 3D object of interest within a 3D scene corresponding to the background image.
9. The computer-implemented method of claim 1 , further comprising:
rendering a 3D scene to produce the background image; and
combining the rendered image of the object of interest and the background image to produce the input image.
10. The computer-implemented method of claim 1 , wherein the task-specific training data defines a location and dimensions of a bounding box enclosing the rendered image of the object of interest.
11. The computer-implemented method of claim 1 , wherein the task-specific training data comprises an object identifier for each pixel that is covered by the rendered image of the object of interest.
12. The computer-implemented method of claim 1 , further comprising receiving a position and orientation in a 3D space for rendering the three-dimensional object of interest.
13. The computer-implemented method of claim 1 , wherein rendering the 3D object of interest comprises applying a texture map to the 3D object of interest.
14. The computer-implemented method of claim 1 , further comprising receiving a position and orientation of a camera for rendering the 3D object of interest.
15. The computer-implemented method of claim 1 , wherein the background image is a synthetic image.
16. The computer-implemented method of claim 1 , wherein the background image is a photorealistic image.
17. The method of claim 1 , wherein the one or more rendered geometric shapes are omitted from the task-specific training data.
18. The method of claim 1 , wherein the object of interest is rendered based on a parameter that is randomly specified to affect an appearance of the object of interest.
19. A system, comprising:
a graphics processing unit (GPU) configured to:
render a three-dimensional (3D) object of interest to produce a rendered image of the object of interest, wherein an input image comprises the rendered image of the object of interest and a background image; and
render one or more 3D geometric shapes to produce one or more rendered geometric shapes, wherein the one or more rendered geometric shapes are included in the input image; and
a memory configured to store a training dataset for training a neural network to detect the object of interest and ignore the one or more rendered geometric shapes, the training dataset including task-specific training data corresponding to the object of interest and the input image as a test pair, wherein the task-specific training data corresponding to the object of interest is computed by a processor coupled to the GPU or the GPU.
20. A non-transitory, computer-readable storage medium storing instructions that, when executed by a processing unit, cause the processing unit to:
render a three-dimensional (3D) object of interest to produce a rendered image of the object of interest, wherein an input image comprises the rendered image of the object of interest and a background image;
render one or more 3D geometric shapes to produce one or more rendered geometric shapes, wherein the one or more rendered geometric shapes are included in the input image;
compute task-specific training data corresponding to the object of interest; and
include the task-specific training data corresponding to the object of interest and the input image as a test pair in a training dataset for training a neural network to detect the object of interest and ignore the one or more rendered geometric shapes.