chahuja / language2pose

Language2Pose: Natural Language Grounded Pose Forecasting
http://chahuja.com/language2pose
MIT License
64 stars 14 forks source link

Question on visualizing the data #10

Closed AloneTogetherY closed 3 years ago

AloneTogetherY commented 3 years ago

Hello and thanks for your work. I was wondering if you could tell me how you were able to animate the joint values from the XML-files, although some coordinates are sometimes missing? Take for example the joint LKx_joint, which only has the x-coordinate. How would you use a 3d matplotlib plot to visualize these joints? I tried examing your code, but I can't find the place where you simply read the XML-file, parse the joint values, and visualize them.

I would be grateful for any tips.

chahuja commented 3 years ago

Hi, The XML files contain joint angles for each joint, but some of these joints only have 2 or 1 degree of motion such as LK (or Left Knee I think). So, here the assumption is that the other joint angles do not change. You can find the skeleton structure here.

The preprocessing class which converts these joint angles to quaternions and then to joint positions (in Euler space) is available here and the commands to preprocess the data is available in the Readme here.

Once the joint positions are extracted it is straightforward to render these animations in matplotlib. The render function is available here.

Hope you find these useful!

AloneTogetherY commented 3 years ago

Thank you :)

AloneTogetherY commented 3 years ago

Another question though...why are the arms of the skeleton hanging behind the its back in most of your animations? Is it the dataset or how it was preprocessed? Thanks again for your time

chahuja commented 3 years ago

For some reason, there was a constant bias in the shoulder joints of the skeleton provided in the dataset. I was not able to rectify it completely, hence the artefacts.