Convolutional neural network for wide-angle camera images

1. An apparatus for applying a convolutional neural network (CNN) to a wide-angle camera image, comprising:

a camera, wherein the camera is to capture a wide-angle image;

a controller to project the wide-angle image onto a three dimensional (3D) surface to obtain an unwrapped image;

a convolution mechanism to perform convolution on the unwrapped image via at least one filter to generate a plurality of features, wherein the wide-angle image projected onto the 3D surface is sampled at a plurality of slices with a same angle of separation; and

a fully connected layer of the CNN to classify the plurality of features generated by the convolution mechanism.

2. The apparatus of claim 1 , wherein the 3D surface is a sphere.

3. The apparatus of claim 1 , wherein projecting the wide-angle image onto the 3D surface comprises multiplying a projection matrix to each real-world coordinate point from the wide-angle image.

4. The apparatus of claim 1 , wherein intrinsic parameters from the camera are used to map the image onto the 3D surface, wherein the intrinsic parameters comprise polynomial coefficients, camera centers, camera resolution, or any combination thereof.

5. The apparatus of claim 1 , wherein a spherical harmonics transformation is applied to the unwrapped image and convolution is performed in a spherical harmonics domain.

6. The apparatus of claim 1 , wherein the fully connected layer is to determine a number of classes used to classify a plurality of features.

7. The apparatus of claim 1 , wherein a polynomial approximation of the camera is used to map the image on the 3D surface.

8. A method for applying a convolutional neural network (CNN) to a wide-angle camera image, comprising:

capturing a wide-angle image;

projecting the wide-angle image onto a unit sphere to obtain an unwrapped image;

convolving the unwrapped image via a plurality of convolution layers to generate a plurality of features, wherein the wide-angle image projected onto the unit sphere is sampled at a plurality of slices of the unit sphere with a same angle of separation; and

classifying the plurality of features extracted by the convolution layers.

9. The method of claim 8 , wherein the wide angle image is represented using spherical harmonics and the convolution is performed using the spherical harmonics.

10. The method of claim 8 , wherein the plurality of convolution layers comprises convolution kernels in a spherical domain.

11. The method of claim 8 , wherein at least one layer of the plurality of convolution layers comprises a linear shift invariant filter.

12. The method of claim 8 , wherein at least one layer of the plurality of convolution layers comprises a rotational symmetric filter.

13. The method of claim 8 , wherein at least one layer of the plurality of convolution layers comprises a spherical Gaussian function.

14. A system for applying a convolutional neural network (CNN) to a wide-angle camera image, comprising

a camera configured to capture a wide-angle image;

a memory that is to store instructions and that is communicatively coupled to the camera; and

a processor communicatively coupled to the camera and the memory, wherein when the processor is to execute the instructions, the processor is to:

project the wide-angle image onto a unit sphere to obtain an unwrapped image;

convolve the unwrapped image via a plurality of convolution layers in a spherical domain to generate a plurality of features, wherein the wide-angle image projected onto the unit sphere is sampled at a plurality of slices of the unit sphere with a same angle of separation; and

classify the plurality of features extracted by the convolution layers.

15. The system of claim 14 , wherein the wide angle image is represented using spherical harmonics and the convolution is performed using the spherical harmonics.

16. The system of claim 14 , wherein a polynomial approximation of a distortion of the wide-angle camera image is used to project the image onto the sphere.

17. The system of claim 14 , wherein at least one layer of the plurality of convolution layers comprises a linear shift invariant filter.

18. The system of claim 14 , wherein projecting the wide-angle image onto the sphere comprises multiplying a projection matrix to each real-world coordinate point from the wide-angle image.

19. The system of claim 14 , wherein intrinsic parameters from a camera are used to project the image onto the sphere, wherein the intrinsic parameters comprise polynomial coefficients, camera centers, camera resolution, or any combination thereof.

20. The system of claim 14 , wherein an inverse spherical harmonics transformation is applied to a result of convolution in the spherical harmonics domain before classifying the plurality features.

21. The system of claim 14 , wherein a fully connected layer of the CNN is used to determine a number of classes used to classify the plurality of features.

22. A tangible, non-transitory, computer-readable medium comprising instructions that, when executed by a processor, direct the processor to:

obtain a wide-angle image;

project the wide-angle image onto a unit sphere to obtain an unwrapped image;

convolve the unwrapped image via a plurality of convolution layers to generate a plurality of features, wherein the wide-angle image projected onto the unit sphere is sampled at a plurality of slices of the unit sphere with a same angle of separation; and

classify the plurality of features extracted by the convolution layers.

23. The computer readable medium of claim 22 , wherein the wide angle image is represented using spherical harmonics and the convolution is performed using the spherical harmonics.

24. The computer readable medium of claim 22 , wherein the plurality of convolution layers comprises convolution kernels in a spherical domain.

25. The computer readable medium of claim 22 , wherein at least one layer of the plurality of convolution layers comprises a linear shift invariant filter.