c3d_ucf101

UCF-101 training demo

Follow these steps to train C3D on UCF-101.

1. Download UCF-101 dataset from UCF-101 website.
2. Unzip the dataset: e.g. unrar x UCF101.rar
3. (Optional) video reader works more stably with extracted frames than directly with video files. Extract frames from UCF-101 videos by revising and running a helper script, ${video-caffe-root}/examples/c3d_ucf101/extract_UCF-101_frames.sh.
4. Change ${video-caffe-root}/examples/c3d_ucf101/c3d_ucf101_{train,test}_split1.txt to correctly point to UCF-101 videos or directories that contain extracted frames.
5. Modify ${video-caffe-root}/examples/c3d_ucf101/c3d_ucf101_train_test.prototxt to your taste or HW specification. Especially batch_size may need to be adjusted for the GPU memory.
6. Run training script: e.g. cd ${video-caffe-root} && examples/c3d_ucf101/train_ucf101.sh
7. After ~7 epochs of training, check if you have about 45% clip accuracy. (See original paper Figure 2 -- ~45% clip accuracy around 6th epoch.)

A typical training will yield the following loss and top-1 accuracy:

Files in this directory

• train_ucf101.sh: a main script to run for training C3D on UCF-101 data
• c3d_ucf101_solver.prototxt: a solver specifications -- SGD parameters, testing parametesr, etc
• c3d_ucf101_test_split1.txt, c3d_ucf101_train_split1.txt: lists of testing/training video clips in ("video directory", "starting frame num", "label") format
• c3d_ucf101_train_test.prototxt: training/testing network model
• ucf101_train_mean.binaryproto: a mean cube calculated from UCF101 training set
• c3d_ucf101_train_loss_accuracy.png: a sample plot of training iteration vs loss and accuracy