Comments (2)
Dear @nlpofwhat,
I have read your anther paper of Moglow (...) It seems to be a autogressive model totally with locomotion?
The StyleGestures paper builds on our earlier MoGlow preprint, which has been accepted for presentation at SIGGRAPH Asia this year. Since the architectures are relatively similar in both works, I would personally argue that both of them are examples of "MoGlow"("motion Glow") architectures, with the main difference being the application: The original MoGlow paper preprint demonstrates that the architecture works for different locomotion-generation tasks, while the StyleGestures follow-up work shows that the architecture can be adapted to gesture synthesis and also adds style control.
So, in summary: My view is that MoGlow is an architecture or a method; locomotion synthesis and speech-driven gesture generation are applications for which we have demonstrated that the method gives good results. MoGlow is thus not an "autogressive model totally with locomotion".
I try to reproduce the results but I can't figure out what the original input data is. For example, you generate gestures by audio then the main input data is audio.
The original MoGlow paper investigates path-based locomotion control from motion-capture data of humans and dogs. For each frame, the main input (the control signal) contains 3 numbers that specify:
- The current forward velocity of the character in a body-centric coordinate system.
- The current sideways velocity of the character in a body-centric coordinate system.
- The current turning velocity of the character.
Together, these 3 numbers act to steer the moving character a bit like driving a car. Over time, the control signal defines a path through space that the root node will follow, along with the character's facing direction at different times. The task of the probabilistic model is then to generate a sequence of poses for the character that represent convincing locomotion along that path.
Like you suggest, this is a very different problem than generating gesture motion using speech acoustics as input features. We think it is a strength of the method that the same basic approach gives such compelling results on such disparate problems.
I can't figure out what the original input data is
If your question is about data availability and coding (e.g., which numbers represent what), the pre-processed locomotion data we used in the paper can be downloaded from: https://kth.box.com/s/quh3rwwl2hedwo32cdg1kq7pff04fjdf. (Edited in 2022: Our university is no longer using Box, invalidating the link originally provided here. Please see the repository readme for the latest files.) Code for training the MoGlow locomotion systems from the article is available in this repository, but may be split off to its own GitHub repo in the near future. Additional information regarding how to use the locomotion training data can be found in issue #1, for instance in this comment as well as this comment and its response. @simonalexanderson knows more about the details of the code and might be able to follow up with any information that I may have missed.
I hope this answers your questions.
from stylegestures.
I've got my answers. Thank you for answering my questions so patiently.
from stylegestures.
Related Issues (20)
- Did you experiment with different batch sizes in training? HOT 2
- Strange results when I train with multiple GPUs. HOT 2
- bvh files with fixed frames
- Difference between time_steps and seqlen? HOT 2
- Possible bug when computing the log-det of Jacobian for affine coupling HOT 2
- About datasets HOT 1
- For the freshmen about Gesture Generation HOT 2
- Questions about the latent random variable Z HOT 2
- The Python version
- About the Cuda version HOT 4
- About the swapaxes for self.x and self.cond HOT 6
- This dataset link doesn't seem to work. HOT 2
- The dataset are inconsistent HOT 10
- Excuse me, where can I find the dataset used by Example3 in readme? HOT 11
- This is the curve when I use two different data sets for training, and the parameters are the same. It can be roughly seen that the loss in Figure 1 will be lower than that in Figure 2, which can indicate that the performance of the first trained model will be better? HOT 6
- How to apply the output file(*.bvh) to 3D model file(*.3ds) HOT 3
- Some questions about the style control HOT 2
- Some questions about the style control HOT 1
- About the dataset HOT 2
- The trained model posture shakes badly. What might be the cause? Is there any way to solve this problem? HOT 2
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from stylegestures.