Normalizing flows for shape analysis.
Part of a master thesis Norwegian University of Science and Technology.
The experiments are placed in the experimets directory. Each of the notebooks corresponds to one of the six experiments in the thesis. The experiments are divided into two categories clustering and interpolation. Each of the categories is then performed on three datasets; test datasets such as moons, frames of human motions, and human motions.
To run the experiments in the notebooks, clone the repository, and install dependencies in requirements.txt. For instance with
pip3 install -r requirements.txt
from the project root. Then most of the notebooks can be run.
To run notebooks with motion capture data the human motion database needs to be installed on your computer. Set up the motion capture database as explained in alexarntzen/signatureshape. This is a copy of paalel/Signatures-in-Shape-Analysis that works with python3.
The videos of motion and frame interpolation are located in the videos directory.
<
Latent space interpolation of human motion (20 interpolating motions of 1 second)