Infuse the 3D CNN model with the assumptions used in optical flow computaitons in a soft way through a special regularization on the filters.
We test the merit of this idea by training ConvNets from scratch on the UCF101 Human Action Recognition data set using Theano. See the report.
- Numpy, Scipy
- Theano (0.7rc1 or later)
- OpenCV (for reading videos)
- LMDB
- Matplotlib (to use analysis tools)