This was forked from https://github.com/toddnguyen/c3d-tensorflow, but debugged/modified significantly to work with existing UCF-101 data set and in dextro environment. It's a WIP for end-to-end training for 8-conv layer C3D network.
Steps to start raining:
- Download UCF-101 dataset from UCF-101 website.
- Unzip the dataset: e.g.
unrar x UCF101.rar
- Extract frames from UCF-101 videos by revising and running a helper script,
bash tools/extract_frames.sh
. - Download pre-trained weights and mean cube files:
bash models/get_weights_and_mean.sh
- Add facebook/C3D python path (requirement for the subsequent steps):
export PYTHONPATH=$PYTHONPATH:/path/to/facebook-c3d/python
- Convert weights:
python convert_weights.py
- Convert a mean cube:
python convert_mean.py