Unsupervised learning of action classes with continuous temporal embedding
Official implementation in python. https://arxiv.org/abs/1904.04189
Two branches: master, global
master: Pipeline for one activity class. Figure 1 in the paper.
global: Proposed pipeline for unsupervised learning with unknown activity classes. Figure 2 in the paper.
conda create --name cte --file requirements.txt
one file per video # rows = # frames in video # columns = dimensionality of frame-wise features
to extract frame-wise features use improved dense trajectories (this step can be substituted by smth else)
Run your own data
see folders TD_utils and test_data and modify files respectively
- Reproduce numbers
- Qualitative results
- Dense trajectorues extaction
- table 1, videovector howto