##Model
This is an implementation of C3D trained on the UCF101 dataset.
The model training script train.py will train an action recogition model from scratch on the UCF101 dataset.
The preprocessor and data loader require OpenCV which can be installed with apt-get.
apt-get install libopencv-dev python-opencv
This model has been tested with OpenCV 2.4.8. If running the preprocessor with a python virtual env then linking the python-opencv library is needed such as by
ln -s /usr/lib/python2.7/dist-packages/cv2.so .venv/lib/python2.7/site-packages/cv2.so
The first step is to preprocess the UCF101 dataset using preprocess.py which splits videos into smaller clips. Preprocessed videos need to be created for both the training and test splits.
python preprocess.py <video_dir> <data_split_file> <class_ind_file> <preprocessed_dir/split_dir>
<video_dir> is the location of the raw video data, <data_split_file> is the name of the training or test split, and <class_ind_file> is a text file containing the mapping from class labels to indices. <preprocessed_dir/split_dir> is the output directory of the preprocessed videos for a given split.
An example of running this program is:
python examples/video-c3d/preprocess.py --video_dir ~/data/UCF-101/ --data_split_file ~/data/ucfTrainTestlist/trainlist01.txt --class_ind_file ~/data/ucfTrainTestlist/classInd.txt --preprocessed_dir ~/data/ucf_preprocessed/train1
python examples/video-c3d/preprocess.py --video_dir ~/data/UCF-101/ --data_split_file ~/data/ucfTrainTestlist/testlist01.txt --class_ind_file ~/data/ucfTrainTestlist/classInd.txt --preprocessed_dir ~/data/ucf_preprocessed/test1
Once the preprocessed video directories are created for both the training and test splits, the model can be trained with the following:
python examples/video-c3d/train.py --data_dir <preprocessed_dir> --batch_size 32 --epochs 18 --save_path UCF101-C3D.p
The data loader will automatically create index files mapping each video clip to its label. Alternatively, an index file can be provided to the data loader.
After the model converges, the demo can be run which predicts the most probable class for each clip and aggregates them into one output video which displays the class labels and their probabilities.
python examples/video-c3d/demo.py --data_dir <preprocessed_dir/split_dir> --model_weights UCF101-C3D.p --class_ind_file ~/data/ucfTrainTestlist/classInd.txt
The weights file for the trained C3D model on the UCF101 action recognition training split 1 can be downloaded from AWS using the following link: trained model weights.
The model achieves 46.3% clip accuracy on the action recognition test split 1 when trained on training split 1.
Learning Spatiotemporal Features with 3D Convolutional Networks
http://arxiv.org/pdf/1412.0767v4.pdf
http://vlg.cs.dartmouth.edu/c3d/
https://github.com/facebook/C3D
Hi! Thanks for your contribution, and I want to know How can I get the C3D model code?