System and method for pose-invariant face alignment

1. A computing system comprising a processing system with at least one processing unit, the processing system being configured to execute a face alignment method that includes:

receiving image data associated with a facial image; and

implementing a neural network to provide a final estimate of parameter data for the facial image based on the image data and an initial estimate of the parameter data, the neural network being configured to generate a feature map based on a current estimate of the parameter data,

wherein:

the neural network is a single convolutional neural network with a plurality of visualization blocks that are connected;

each visualization block includes at least a visualization layer, convolutional layers, and fully connected layers; and

the parameter data includes head pose data and face shape data.

2. The computing system of claim 1 , further comprising:

a face detection module configured to at least (i) receive an image, (ii) identify the facial image within the image, and (iii) provide the image data relating to the facial image to the processing system,

wherein the facial image provides a facial view that is within a range of a frontal view to a profile view.

3. The computing system of claim 1 , wherein the face shape data includes facial identity parameters and facial expression parameters.

4. The computing system of claim 1 , wherein the single convolutional neural network includes end-to-end training for model fitting.

5. The computing system of claim 1 , wherein each visualization layer is based on surface normals of a 3D face model and encodes a relative head pose between a face and a camera.

6. The computing system of claim 1 , wherein each visualization layer utilizes a mask to differentiate between pixels in different parts of the facial image and make pixel values of visualized images similar across different head poses.

7. A computer-implemented method for face alignment comprising:

receiving image data associated with a facial image; and

implementing a neural network to provide a final estimate of parameter data for the facial image based on the image data and an initial estimate of the parameter data, the neural network including at least one visualization layer that is configured to generate a feature map based on a current estimate of the parameter data,

wherein:

each visualization layer utilizes a mask to differentiate between pixels in different parts of the facial image and make pixel values of visualized images similar across different head poses; and

the parameter data includes head pose data and face shape data.

8. The computer-implemented method of claim 7 , further comprising:

performing face detection on an image that includes (i) receiving the image, (ii) identifying the facial image within the image, and (iii) providing the image data relating to the facial image to the processing system,

wherein the facial image provides a facial view that is within a range of a frontal view to a profile view.

9. The computer-implemented method of claim 7 , wherein the face shape data includes facial identity parameters and facial expression parameters.

10. The computer-implemented method of claim 7 , wherein the neural network is a single convolutional neural network with end-to-end training for model fitting.

11. The computer-implemented method of claim 7 , wherein:

the neural network is a single convolutional neural network with a plurality of visualization blocks that are connected; and

each visualization block includes at least the visualization layer, convolutional layers, and fully connected layers.

12. The computer-implemented method of claim 7 , wherein each visualization layer is based on surface normals of a 3D face model and encodes a relative head pose between a face and a camera.

13. Non-transitory computer-readable media comprising at least computer-readable data that, when executed by a processing system with at least one processing unit, performs a face alignment method that includes:

receiving image data associated with a facial image; and

implementing a neural network to provide a final estimate of parameter data for the facial image based on the image data and an initial estimate of the parameter data, the neural network being configured to generate a feature map based on a current estimate of the parameter data,

wherein:

the neural network is a single convolutional neural network with a plurality of visualization blocks that are connected;

each visualization block includes at least one visualization layer, convolutional layers, and fully connected layers; and

the parameter data includes head pose data and face shape data.

14. The non-transitory computer-readable media of claim 13 , wherein the face shape data includes facial identity parameters and facial expression parameters.

15. The non-transitory computer-readable media of claim 13 , wherein the single convolutional neural network includes end-to-end training for model fitting.

16. The non-transitory computer-readable media of claim 13 , wherein each visualization layer is based on surface normals of a 3D face model and encodes a relative head pose between a face and a camera.

17. The non-transitory computer-readable media of claim 13 , wherein each visualization layer utilizes a mask to differentiate between pixels in different parts of the facial image and make pixel values of visualized images similar across different head poses.