Deep learning-based facial animation for head-mounted display
There is disclosed a system and method for training a set of expression and neutral convolutional neural networks using a single performance mapped to a set of known phonemes and visemes in the form predetermined sentences and facial expressions. Then, subsequent training of the convolutional neural networks can occur using temporal data derived from audio data within the original performance mapped to a set of professionally-created three dimensional animations. Thereafter, with sufficient training, the expression and neutral convolutional neural networks can generate facial animations from facial image data in real-time without individual specific training.
1. A method of real-time facial animation comprising:
receiving video and audio of at least a portion of a performer's face from at least one performance of a predetermined set of phonemes and facial expressions;
generating a three-dimensional model of facial animations corresponding to the video and audio from the at least one performance relying in part upon temporal data for the at least one performance derived from the audio;
capturing a new performance of video and audio for a new performer's face including the predetermined set of phonemes and facial expressions;
applying a first convolutional neural network to derive a set of expression blendshape weights for each frame of the new performance of video, the expression blendshape weights also based upon at least one frame of video preceding a frame for which expression weights are sought;
applying a second convolutional neural network to each frame of the new performance of video to derive neutral facial weights;
applying a mapping function to the new performance to create a facial animation blending the results of the first convolutional neural network and the second convolutional neural network using the temporal data to match the facial animation to the new performance.
2. The method of claim 1 further comprising applying both the first and second convolutional neural networks to an unrelated video performance with a blendweight selected by the first and second convolutional neural networks to generate a facial animation that most-closely matches the unrelated video performance.
3. The method of claim 1 wherein the mapping function relies upon a blendshape weight vector for the first convolutional neural network and the second convolutional neural network that causes the mapping function to best corresponds to a face shape the new performance.
4. The method of claim 1 wherein the first convolutional neural network generates a three-dimensional map of a facial animation that best corresponds to the new performance.
5. The method of claim 1 wherein the at least one frame is two video frames of the new performance.
6. The method of claim 1 wherein the at least one performance is more than five performances.
7. The method of claim 1 wherein the generating the three-dimensional model is computer-assisted using temporal data derived from the at least one performance audio so as to temporally scale the at least one performance so as to automatically generate estimates of appropriate facial animations for each of the at least one performance.
8. Apparatus comprising a storage medium storing a program having instructions which when executed by a processor will cause the processor to:
receive video and audio of at least a portion of a performer's face from at least one performance of a predetermined set of phonemes and facial expressions;
generate a three-dimensional model of facial animations corresponding to the video and audio from the at least one performance relying in part upon temporal data for the at least one performance derived from the audio;
capture a new performance of video and audio for a new performer's face including the predetermined set of phonemes and facial expressions;
apply a first convolutional neural network to derive a set of expression blendshape weights for each frame of the new performance of video, the expression blendshape weights also based upon at least one frame of video preceding a frame for which expression weights are sought;
apply a second convolutional neural network to each frame of the new performance of video to derive neutral facial weights;
apply a mapping function to the new performance to create a facial animation blending the results of the first convolutional neural network and the second convolutional neural network using the temporal data to match the facial animation to the new performance.
9. The apparatus of claim 8 wherein the processor is further instructed to apply both the first and second convolutional neural networks to an unrelated video performance with a blendweight selected by the first and second convolutional neural networks to generate a facial animation that most-closely matches the unrelated video performance.
10. The apparatus of claim 8 wherein the mapping function relies upon a blendshape weight vector for the first convolutional neural network and the second convolutional neural network that causes the mapping function to best corresponds to a face shape the new performance.
11. The apparatus of claim 8 wherein the first convolutional neural network generates a three-dimensional map of a facial animation that best corresponds to the new performance.
12. The apparatus of claim 8 wherein the at least one frame is two video frames of the new performance.
13. The apparatus of claim 8 wherein the at least one performance is more than five performances.
14. The apparatus of claim 8 wherein generation of the three-dimensional model is computer-assisted using temporal data derived from the at least one performance audio so as to temporally scale the at least one performance so as to automatically generate estimates of appropriate facial animations for each of the at least one performance.
15. The apparatus of claim 8 further comprising:
the processor;
a memory; and
wherein the processor and the memory comprise circuits and software for performing the instructions on the storage medium.
16. A system for real-time facial animation comprising:
at least one camera and microphone for:
capturing video and audio of at least a portion of a performer's face from at least one performance of a predetermined set of phonemes and facial expressions;
capturing a new performance of video and audio for a new performer's face including the predetermined set of phonemes and facial expressions;
a processor for:
generating a three-dimensional model of facial animations corresponding to the video and audio from the at least one performance relying in part upon temporal data for the at least one performance derived from the audio;
applying a first convolutional neural network to derive a set of expression blendshape weights for each frame of the new performance of video, the expression blendshape weights also based upon at least one frame of video preceding a frame for which expression weights are sought;
applying a second convolutional neural network to each frame of the new performance of video to derive neutral facial weights;
applying a mapping function to the new performance to create a facial animation blending the results of the first convolutional neural network and the second convolutional neural network using the temporal data to match the facial animation to the new performance.
17. The system of claim 16 wherein the processor is further for applying both the first and second convolutional neural networks to an unrelated video performance with a blendweight selected by the first and second convolutional neural networks to generate a facial animation that most-closely matches the unrelated video performance.
18. The system of claim 16 wherein the mapping function relies upon a blendshape weight vector for the first convolutional neural network and the second convolutional neural network that causes the mapping function to best corresponds to a face shape the new performance.
19. The system of claim 16 wherein the first convolutional neural network generates a three-dimensional map of a facial animation that best corresponds to the new performance.
20. The system of claim 16 wherein generation of the three-dimensional model is computer-assisted using temporal data derived from the at least one performance audio so as to temporally scale the at least one performance so as to automatically generate estimates of appropriate facial animations for each of the at least one performance.