Robust facial animation from video and audio
Implementations described herein relate to methods, systems, and computer-readable media to generate animations for a 3D avatar from input video and audio captured at a client device. A camera may capture video of a face while a trained face detection model and a trained regression model output a set of video FACS weights, head poses, and facial landmarks to be translated into the animations of the 3D avatar. Additionally, a microphone may capture audio uttered by a user while a trained facial movement detection model and a trained regression model output a set of audio FACS weights. Additionally, a blending term is provided for identification of lapses in audio. A modularity mixing component fuses the video FACS weights and the audio FACS weights based on the blending term to create final FACS weights for animating the user's avatar, a character rig, or another animation-capable construct.
1 . A computer-implemented method, comprising:
receiving input video frames;
receiving input audio frames and a blending term, wherein the input audio frames include audio associated with the input video frames;
obtaining video facial action coding system (FACS) weights from a first trained machine learning model based on the input video frames, wherein the first trained machine learning model comprises at least one encoder and at least three task-specific decoders, and wherein:
a first task-specific decoder of the at least three task-specific decoders is configured to output a predicted headpose,
a second task-specific decoder of the at least three task-specific decoders is configured to output a probability that a face is visible in an input video frame of the received input video frames, and
a third task-specific decoder of the at least three task-specific decoders is configured to output facial landmarks;
obtaining audio FACS weights from a second trained machine learning model based on the input audio frames;
combining the video FACS weights and the audio FACS weights to obtain final FACS weights, wherein the combining is based at least in part on the blending term; and
outputting the final FACS weights to drive facial animation of a three-dimensional (3D) model.
2 . The computer-implemented method of claim 1 , wherein the first trained machine learning model and the second trained machine learning model are trained in a two-stage semi-supervised training process.
3 . The computer-implemented method of claim 1 , wherein at least one of the at least three task-specific decoders comprises causal convolution layers applied over a time dimension.
4 . The computer-implemented method of claim 1 , wherein the second trained machine learning model comprises at least one encoder and at least two task-specific decoders.
5 . The computer-implemented method of claim 4 , wherein a first task-specific decoder of the at least two task-specific decoders of the second trained machine learning model is configured to output the audio FACS weights and a second task-specific decoder of the at least two task-specific decoders of the second trained machine learning model is configured to output the blending term.
6 . A system, comprising:
a memory with instructions stored thereon; and
a processing device, coupled to the memory, wherein the processing device is configured to access the memory and execute the instructions, and wherein the instructions cause the processing device to perform operations comprising:
receiving input video frames;
receiving input audio frames and a blending term, wherein the input audio frames include audio associated with the input video frames;
obtaining video facial action coding system (FACS) weights from a first trained machine learning model based on the input video frames, wherein the first trained machine learning model comprises at least one encoder and at least three task-specific decoders, and wherein:
a first task-specific decoder of the at least three task-specific decoders is configured to output a predicted headpose,
a second task-specific decoder of the at least three task-specific decoders is configured to output a probability that a face is visible in an input video frame of the received input video frames, and
a third task-specific decoder of the at least three task-specific decoders is configured to output facial landmarks;
obtaining audio FACS weights from a second trained machine learning model based on the input audio frames;
combining the video FACS weights and the audio FACS weights to obtain final FACS weights, wherein the combining is based at least in part on the blending term; and
outputting the final FACS weights to drive facial animation of a three-dimensional (3D) model.
7 . The system of claim 6 , wherein the first trained machine learning model and the second trained machine learning model are trained in a two-stage semi-supervised training process.
8 . The system of claim 6 , wherein at least one of the at least three task-specific decoders comprises causal convolution layers applied over a time dimension.
9 . The system of claim 6 , wherein the second trained machine learning model comprises at least one encoder and at least two task-specific decoders.
10 . The system of claim 9 , wherein a first task-specific decoder of the at least two task-specific decoders of the second trained machine learning model is configured to output the audio FACS weights and a second task-specific decoder of the at least two task-specific decoders of the second trained machine learning model is configured to output the blending term.
11 . A non-transitory computer-readable medium with instructions stored thereon that, responsive to execution by a processing device, causes the processing device to perform operations comprising:
receiving input video frames;
receiving input audio frames and a blending term, wherein the input audio frames include audio associated with the input video frames;
obtaining video facial action coding system (FACS) weights from a first trained machine learning model based on the input video frames, wherein the first trained machine learning model comprises at least one encoder and at least three task-specific decoders, and wherein:
a first task-specific decoder of the at least three task-specific decoders is configured to output a predicted headpose,
a second task-specific decoder of the at least three task-specific decoders is configured to output a probability that a face is visible in an input video frame of the received input video frames, and
a third task-specific decoder of the at least three task-specific decoders is configured to output facial landmarks;
obtaining audio FACS weights from a second trained machine learning model based on the input audio frames;
combining the video FACS weights and the audio FACS weights to obtain final FACS weights, wherein the combining is based at least in part on the blending term; and
outputting the final FACS weights to drive facial animation of a three-dimensional (3D) model.
12 . The non-transitory computer-readable medium of claim 11 , wherein the first trained machine learning model and the second trained machine learning model are trained in a two-stage semi-supervised training process.
13 . The non-transitory computer-readable medium of claim 11 , wherein at least one of the at least three task-specific decoders comprises causal convolution layers applied over a time dimension.
14 . The non-transitory computer-readable medium of claim 11 , wherein the second trained machine learning model comprises at least one encoder and at least two task-specific decoders, and wherein a first task-specific decoder of the at least two task-specific decoders of the second trained machine learning model is configured to output the audio FACS weights and a second task-specific decoder of the at least two task-specific decoders of the second trained machine learning model is configured to output the blending term.