Skip to content

Video-friendly caffe -- comes with a recent version of Caffe (June 2016), a video reader, 3D(ND) pooling layer, and an example training script for C3D network and UCF-101 data

License

Notifications You must be signed in to change notification settings

ergysr/video-caffe

 
 

Repository files navigation

Video-Caffe: Caffe with C3D implementation and video reader

Build Status

This is 3-D Convolution (C3D) and video reader implementation in the latest Caffe (Apr 2016). The original Facebook C3D implementation is branched out from Caffe on July 17, 2014 with git commit b80fc86, and has not been rebased with the original Caffe, hence missing out quite a few new features in the lastest Caffe. I therefore pulled in C3D concept and an accompanying video reader and applied to the latest Caffe, and will try to rebase this repo with the upstream whenever there is a new important feature. This video-caffe is rebased on df412a, on June 2, 2016. Please reach me for any feedback or question.

Check out the original Caffe readme for Caffe-specific information.

Requirements

In addition to prerequisites for Caffe, video-caffe depends on cuDNN. It is known to work with CuDNN v4 and v5(RC), but it may need some tweaks to build with v3.

  • If you use "make" to build make sure Makefile.config point to the right paths for CUDA and CuDNN.
  • If you use "cmake" to build, double-check CUDNN_INCLUDE and CUDNN_LIBRARY. You may want to cmake with something like cmake -DCUDNN_INCLUDE="/your/path/to/include" -DCUDNN_LIBRARY="/your/path/to/lib" ${video-caffe-root}.

Building video-caffe

In a nutshell, key steps to build video-caffe are:

  1. git clone [email protected]:chuckcho/video-caffe.git
  2. cd video-caffe
  3. mkdir build && cd build
  4. cmake ..
  5. Make sure CUDA and CuDNN are detected and their paths are correct.
  6. make all
  7. make install
  8. (optional) make runtest

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 have to be reduced for the GPU memory. Should run fine as is with 6GB GPU memory.
  6. Run training script: e.g. cd ${video-caffe-root} && examples/c3d_ucf101/train_ucf101.sh
  7. Sit back and enjoy.

License and Citation

Caffe is released under the BSD 2-Clause license.

About

Video-friendly caffe -- comes with a recent version of Caffe (June 2016), a video reader, 3D(ND) pooling layer, and an example training script for C3D network and UCF-101 data

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • C++ 80.1%
  • Python 7.9%
  • Cuda 5.8%
  • CMake 2.7%
  • Protocol Buffer 1.6%
  • MATLAB 0.9%
  • Other 1.0%