this project implements action recognition algorithm proposed in C3D: Generic Features for Video Analysis with esimator of Tensorflow
c3d is a convolutional neural network classifying sports video clips. it is widely used as an infrastructure of latter action recognition neural networks.
a trained model is provided on baidu cloud at
if you want to train yourself, you need UCF101 dataset. download it and extract the directory. set the root directory in create_dataset.py. then create a tfrecord format dataset with command
python3 create_dataset.pythe program will generate a trainset and a testset tfrecord file.
start the training by exeucting
python3 train_c3d.pymoniter the training process by tensorboard and stop the training when the accuracy reaches 82% which is the best accuracy c3d can reach.
test on single video clip by modifying the video path in ActionRecognition.py and run with
python3 ActionRecognition.pya readable label will be printed on the video.
run following command to get serving model
python3 convert_model.py