Detecting one or more objects in an image, or sequence of images, and determining a category and one or more descriptors for each of the one or more objects, generating synthetic training data, and training a neural network with the synthetic training data
In an embodiment, an image-capture system, includes an image-capture device and computing circuitry. The image-capture device is configured to capture an image of a region of space that includes an object. And the computing circuitry is coupled to the image-capture device and is configured to detect a representation of the object in the image, to determine a representation of a boundary of the detected representation, to provide image information corresponding to the detected representation to an image-analysis system, to receive, from the image-analysis system, an identifier of a category to which the object belongs, and a descriptor of the object, and to generate a representation of a list that includes the identifier and the descriptor.
1. A method for generating at least one synthetic training image, the method comprising:
receiving an image;
receiving at least one three-dimensional representation of an object;
implementing, using a rendering engine, at least one instruction for altering at least one parameter of the at least one three-dimensional representation of the object;
generating at least one synthetic training image that includes at least the altered parameter of the three-dimensional representation of the object;
associating metadata with the at least one synthetic training image;
populating the image with the synthetic training image that includes at least the altered parameter of the three-dimensional representation of the object;
generating a second synthetic training image using at least the populated image; and
storing the associated metadata and the second synthetic training image in memory.
2. The method of claim 1 , wherein the three-dimensional representation of an object is at least one of a computer-aided-design (CAD) electronic file, a point cloud, or a portion of a first synthetic training image.
3. The method of claim 1 , wherein altering at least one parameter comprises adjusting at least one lighting condition of the three-dimensional representation of the object.
4. The method of claim 1 , wherein altering at least one parameter comprises adjusting a sensor parameter.
5. The method of claim 4 , wherein the adjusting the sensor parameter includes altering at least one of a light, sound, infrared, or thermal parameter.
6. The method of claim 1 , wherein altering at least one parameter comprises adjusting an orientation of the three-dimensional representation of the object.
7. The method of claim 5 , further comprising creating a two-dimensional representation of the three-dimensional representation of the object having an adjusted orientation.
8. The method of claim 4 , wherein the adjusting the sensor parameter replicates a camera condition, the camera condition including at least one of a camera angle, a resolution, a bit depth, a dynamic range, a camera zoom, f-stop, color balance, contrast, focus, and color temperature.
9. The method of claim 1 , wherein the at least one instruction for altering the at least one parameter alters one or more of a space between the object and a second object, a number of objects within a space, an object orientation, a lighting condition, a texture of the object, and a feature of the object.
10. The method of claim 1 , further comprising:
a. storing the associated metadata and the at least one synthetic training image using at least the populated image in memory.
11. The method of claim 1 , wherein the received image is at least one of a still image, a sequence of images, a satellite image, or a high-dynamic-range (HDR) image.
12. The method of claim 1 , wherein generating the second synthetic training image further comprises:
a. receiving at least one sensor condition; and
b. adjusting at least one image parameter based at least in part on the received at least one sensor condition.
13. The method of claim 1 , further comprising:
a. implementing at least one instruction for altering at least one parameter of the at least one three-dimensional representation of the object of the at least one synthetic training image to generate at least one second synthetic training image;
b. associating metadata with the at least one second synthetic training image;
c. storing the associated metadata and the at least one second synthetic training image in memory; and
d. training at least one neural network to recognize at least one portion of the object using the at least one synthetic training image and the at least one second synthetic training image.
14. The method of claim 13 , wherein the at least one altered parameter is a texture of the first synthetic training image.
15. The method of claim 13 , wherein the at least one altered parameter includes at least one lighting condition of the first synthetic training image.
16. The method of claim 13 , wherein the at least one altered parameter includes at least one camera condition of the synthetic training image.
17. The method of claim 13 , wherein implementing the at least one instruction generates a substantially random set of statistically independent parameters of the object.
18. The method of claim 17 , wherein the at least one instruction includes at least a user input.
19. The method of claim 13 , wherein generating the at least one second synthetic training image includes applying a texture map to the object.
20. The method of claim 1 , wherein the associated metadata further comprises a category, a common category, one or more descriptors, an object identifier, a list of objects within the first synthetic training image, a pixel map, or an object pairing.
21. A method for randomly generating a plurality of synthetic training images, the method comprising:
receiving an image;
receiving at least one three-dimensional representation of an object;
implementing at least one instruction for altering one or more parameters of the at least one three-dimensional representation of the object;
generating at least one synthetic training image that includes at least the altered parameter of the three-dimensional representation of the object, wherein the at least one synthetic training image includes a statistically independent spread of the one or more parameters amongst the plurality of synthetic training images;
associating metadata with the at least one synthetic training image; populating the image with the synthetic training image that includes at least the altered parameter of the three-dimensional representation of the object; and
generating a second synthetic training image using at least the populated image receiving at least one three-dimensional representation of an object; and
storing the associated metadata and the second synthetic training image in memory.
22. The method of claim 21 , wherein the associating metadata further comprises extracting the metadata from at least one of a list, embedded metadata within a file containing the at least one three-dimensional representation of the object, or metadata from a server.
23. The method of claim 21 , further comprising:
receiving image data relating to the at least one three-dimensional representation of the object;
altering at least one parameter of the image data relating to the at least one three-dimensional representation of the object;
generating at least one synthetic image that includes at least the altered parameter of the image data relating to the at least one object; and
associating metadata with the at least one synthetic image.
24. The method of claim 21 , wherein the received image is at least one of a still image, a sequence of images, a satellite image, or a high-dynamic-range (HDR) image.
25. The method of claim 21 , further comprising receiving image information including one or more of: a category of at least one object within the image, at least one descriptor of the at least one object within the image, an identifier, or sensor information of the received image.
26. The method of claim 21 , further comprising training at least one neural network to recognize at least one portion of the object using at least the first synthetic training image.
27. The method of claim 21 , wherein the at least one neural network further performs a grouping action of at least one object in a plurality of synthetic training images into at least one hierarchal cluster.