Deep-learning motion priors for full-body performance capture in real-time
Training data from multiple types of sensors and captured in previous capture sessions can be fused within a physics-based tracking framework to train motion priors using different deep learning techniques, such as convolutional neural networks (CNN) and Recurrent Temporal Restricted Boltzmann Machines (RTRBMs). In embodiments employing one or more CNNs, two streams of filters can be used. In those embodiments, one stream of the filters can be used to learn the temporal information and the other stream of the filters can be used to learn spatial information. In embodiments employing one or more RTRBMs, all visible nodes of the RTRBMs can be clamped with values obtained from the training data or data synthesized from the training data. In cases where sensor data is unavailable, the input nodes may be unclamped and the one or more RTRBMs can generate the missing sensor data.
1. A method for motion capture, the method being implemented by a processor configured to execute machine-readable instructions, the method comprising:
obtaining a deep learning model, wherein the deep learning model includes a convolutional neural network (CNN);
obtaining a set of training data and training the deep learning model using the training data to generate a motion model, the training data including temporal and spatial information regarding one or more actors' motion captured previously, wherein the training of the deep learning model includes:
receiving the spatial information in a first stream to learn a spatial relationship of the one or more actors' motion captured previously using a first set of one or more filters;
receiving the temporal information in a second stream to learn a temporal relationship of the deep learning model using a second set of one or more filters, wherein the second stream is received independent of the first stream;
mapping the set of training data to a ground truth skeleton pose using a fully connected feedforward neural network;
receiving, from one or more motion capture sensors, real-time motion data for the one or more actor's motion;
identifying missing motion information in the received real-time motion data introduced by the one or more motion capture sensors; and
estimating the missing motion information regarding the actor's motion based on the received real-time motion data using the deep learning model with inverse dynamics to estimate a reference pose for a frame and combining the received real-time motion data with the estimated reference pose to solve for orientation and position constraints.
2. The method of claim 1 , wherein the training of the motion model further includes connecting the first and second streams using a recurrent neural net (RNN).
3. The method of claim 2 , wherein the training of the motion model further includes training the motion model using a stochastic gradient decent with Rectified Linear Unit (RELU) activation.
4. The method of claim 1 , wherein the deep learning model includes a recurrent temporal restricted Boltzmann machine (RTRBM), and wherein training of the motion model includes receiving raw sensor data and presenting the sensor data to a network in a last visible layer of the RTRBM.
5. The method of claim 4 , wherein training of the motion model includes synthesizing the raw sensor data and clamping all visible nodes with at least one of the raw sensor data and the synthesized sensor data.
6. A system for motion capture, the system comprising one or more of a processor configured to execute machine-readable instructions such that when the machine-readable instructions are executed, the process is caused to perform:
obtaining a deep learning model, wherein the deep learning model includes a convolutional neural network (CNN);
obtaining a set of training data and training the deep learning model using the training data to generate a motion model, the training data including temporal and spatial information regarding one or more actors' motion captured previously, wherein the training of the deep learning model includes:
receiving the spatial information in a first stream to learn a spatial relationship of the one or more actors' motion captured previously using a first set of one or more filters;
receiving the temporal information in a second stream to learn a temporal relationship of the deep learning model using a second set of one or more filters, wherein the second stream is received independent of the first stream;
mapping the set of training data to a ground truth skeleton pose using a fully connected feedforward neural network;
receiving, from one or more motion capture sensors, real-time motion data for the one or more actor's motion;
identifying missing motion information in the received real-time motion data introduced by the one or more motion capture sensors; and
estimating the missing motion information regarding the actor's motion based on the received real-time motion data using the deep learning model with inverse dynamics to estimate a reference pose for a frame and combining the received real-time motion data with the estimated reference pose to solve for orientation and position constraints.
7. The system of claim 6 , wherein the training of the motion model further includes connecting the first and second streams using a recurrent neural net (RNN).
8. The system of claim 7 , wherein the training of the motion model further includes training the motion model using a stochastic gradient decent with Rectified Linear Unit (RELU) activation.
9. The system of claim 6 , wherein the deep learning model includes a recurrent temporal restricted Boltzmann machine (RTRBM), and wherein training of the motion model includes receiving raw sensor data and presenting the sensor data to a network in a last visible layer of the RTRBM.
10. The system of claim 9 , wherein training of the motion model includes synthesizing the raw sensor data and clamping all visible nodes with the raw sensor data and/or the synthesized sensor data.
11. A computer-readable medium storing a plurality of instructions that, when executed by one or more processors of a computing device, cause the one or more processors to perform operations comprising:
obtaining a deep learning model, wherein the deep learning model includes a convolutional neural network (CNN);
obtaining a set of training data and training the deep learning model using the training data to generate a motion model, the training data including temporal and spatial information regarding one or more actors' motion captured previously, wherein the training of the deep learning model includes:
receiving the spatial information in a first stream to learn a spatial relationship of the one or more actors' motion captured previously using a first set of one or more filters;
receiving the temporal information in a second stream to learn a temporal relationship of the deep learning model using a second set of one or more filters, wherein the second stream is received independent of the first stream;
mapping the set of training data to a ground truth skeleton pose using a fully connected feedforward neural network;
receiving, from one or more motion capture sensors, real-time motion data for the one or more actor's motion;
identifying missing motion information in the received real-time motion data introduced by the one or more motion capture sensors; and
estimating the missing motion information regarding the actor's motion based on the received real-time motion data using the deep learning model with inverse dynamics to estimate a reference pose for a frame and combining the received real-time motion data with the estimated reference pose to solve for orientation and position constraints.