Hidden Two-Stream Convolutional Networks for Action Recognition

This is the Caffe implementation of the "Hidden Two-Stream Convolutional Networks for Action Recognition". You can refer to paper for more details at Arxiv.

Dependencies

OpenCV 3 (Installation can be refered here)

Tested on Ubuntu 16.04 with Titan X GPU, CUDNN 5.1

Compiling

To get started, first compile caffe, by configuring a

"Makefile.config"

then make with

$ make -j 6 all

Training

(this assumes you compiled the code sucessfully)

Here, we take UCF101 split 1 as an example.

First, go to folder,

cd models/ucf101_split1_unsup_end

Then change the FRAME_PATH in train_rgb_split1.txt and val_rgb_split1.txt to where you store the extracted video frames,

/FRAME_PATH/WallPushups/v_WallPushups_g21_c06 111 98

This follows the format as in TSN. 111 indicates the number of frames of that video clip, and 98 represents the action label. For more details about how to construct file list for training and validation, we refer you to here.

Then you need to download the initialization models (pre-trained temporal stream CNN stacked upon pre-trained MotionNet),

UCF101 split1

Then tune the parameters in end_train_val.prototxt and end_solver.prototxt as you need, or leave as it is.

Finally, you can simply run

../../build/tools/caffe train -solver=end_solver.prototxt -weights=ucf101_split1_vgg16_init.caffemodel

NOTE: It is highly likely that you may get better performance than us if you carefully tune the hyper-params such as loss weights, learning rate etc.

Testing

(this assumes you compiled the code sucessfully)

First, download our trained models:

UCF101 split 1 UCF101 split 2 UCF101 split 3

HMDB51 split 1 HMDB51 split 2 HMDB51 split 3

Then go to this folder

cd models/ucf101_split1_unsup_end/eval_ucf101

Then run

python demo_hidden.py

But maybe you need to set paths correctly in demo_hidden.py before you run it, like model_def_file and model_file. And also change the FRAME_PATH in testlist01_with_labels.txt.

MotionNet

The training and testing code of MotionNet is in folder

cd models/multiframe_MotionNet

The pretraied model can be downloaded at MotionNet.

Misc

There is a chance that you may get a little bit higher or lower accuracy on UCF101 and HMDB51 than the numbers reported in our paper, even using our provided trained models. This is normal because your extracted video frames may not be the same as ours, and the quality of image has an impact on the final performance. Thus, no need to raise an issue unless the performance gap is large, e.g. larger than 1%.
Since there are so many losses to compute, you may encounter model divergence in the very beginning of the training. You can simply reduce learning rate first to get a good initialization, and then back on track. Or you just rerun training several times.

TODO

Add other initialization models for the rest splits of UCF101 and HMDB51
Experiment on large-scale action datasets, like Sports-1M and Kinetics

License and Citation

Please cite this paper in your publications if you use this code for your research:

@article{hidden_ar_zhu_2017,
  title={{Hidden Two-Stream Convolutional Networks for Action Recognition}},
  author={Yi Zhu and Zhenzhong Lan and Shawn Newsam and Alexander G. Hauptmann},
  journal={arXiv preprint arXiv:1704.00389},
  year={2017}
}

Related Projects

GuidedNet: Guided Optical Flow Learning

Acknowledgement

The code base is borrowed from TSN, DispNet and UnsupFlownet. Thanks for open sourcing the code.

jqqqqqqqqqq/Hidden-Two-Stream