Skip to content
Official implementation of the paper: Unsupervised learning of action classes with continuous temporal embedding
Branch: master
Clone or download
Latest commit 07ef997 Apr 16, 2019
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
BF_utils need to compute new numbers for global Apr 5, 2019
FS_utils
TD_utils
YTI_utils
eval_utils
models
probabilistic_utils first commit Jan 8, 2019
test_data rm some logs Apr 16, 2019
utils update howto 0.1 + simplify usage Apr 16, 2019
viterbi_utils update howto 0.1 + simplify usage Apr 16, 2019
README.md
corpus.py update howto 0.1 + simplify usage Apr 16, 2019
pipeline.py update howto 0.1 + simplify usage Apr 16, 2019
requirements.txt
video.py update howto 0.1 + simplify usage Apr 16, 2019

README.md

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.

Create environment
conda create --name cte --file requirements.txt
Input files
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

TODO descriptions
  • Evaluation
  • Reproduce numbers
  • Qualitative results
  • Dense trajectorues extaction
  • table 1, videovector howto
You can’t perform that action at this time.