Information

This package contains a matlab implementation of Pose-based CNN (P-CNN) algorithm described in [1]. It includes pre-trained CNN appearance vgg-f model [2], a matlab version of the flow model of [3] and the optical flow implementation of [4]. CNN implementation uses the MatConvNet library [5]. The project webpage is http://www.di.ens.fr/willow/research/p-cnn/ .

To run this package:

• Prepare/download CNN models and data examples by running init.sh file from the P-CNN folder.
• This package compiles MatConvNet [5] in "CPU mode". To speed up computation you may want to enable GPU support (much faster). To help you, we provide the my_build.m file in the matconvnet-1.0-beta11 folder that you can modify.
• You may want to recompile Brox optical flow 2004 [4] (download sources).

demo.m

An example of P-CNN computation is given in this package. It computes P-CNN for few videos of the JHMDB dataset [6] (for 2 different splits) using pose ground truth annotations. The reproduce_ICCV15_results command reproduces the P-CNN results reported in [1]. Because we wanted to provide a "full matlab code", we converted all the code to matlab resulting to a slightly different result (-0.9% accuracy) from the published version due to the switch of the CNN package and retraining.

The provided algorithm takes as input the frames of a video and their corresponding pose joints (from ground truth annotation or from your favorite pose detector). There is a demo.m file in the package that you should be able to run.

Datasets

Two datasets have been used in our ICCV'15 paper:

• JHMDB [6]: as explained above, the demo.m file shows how to use P-CNN with this dataset. The dataset and the ground-truth joint positions can be download here.
• MPII Cooking Activities [7]: You can download the dataset and the estimated joint positions we computed for our experiments. Note that, in MPII Cooking Activities, we do not use the same parameters as for JHMDB (e.g there is no full body part). Then, you have to modify the following parameters in the demo.m file:

param.lhandposition=11;
param.rhandposition=6;
param.upbodypositions=1:13;
param.lside = 120 ;

and in compute_pcnn_features.m:

param.partids = [1 2 3 4] ; % don't use full body part

Cite

If you use this package, please cite:

@inproceedings{cheronICCV15,
TITLE = {{P-CNN: Pose-based CNN Features for Action Recognition}},
AUTHOR = {Ch{'e}ron, Guilhem and Laptev, Ivan and Schmid, Cordelia},
BOOKTITLE = {ICCV},
YEAR = {2015},
}

References

[1] G. Chéron, I. Laptev, C. Schmid. P-CNN: Pose-based CNN Features for Action Recognition. ICCV 2015.

[2] K. Chatfield, K. Simonyan, A. Vedaldi, and A. Zisserman. Return of the devil in the details: Delving deep into convolutional nets. BMVC 2014.

[3] G. Gkioxari and J. Malik. Finding action tubes. CVPR 2015. ACM 2015.

[4] T. Brox, A. Bruhn, N. Papenberg, and J. Weickert. High accuracy optical flow estimation based on a theory for warping. ECCV 2004.

[5] A. Vedaldi and K. Lenc. MatConvNet - Convolutional Neural Networks for MATLAB.

[6] H. Jhuang, J. Gall, S. Zuffi, C. Schmid, and M. J. Black. Towards understanding action recognition. ICCV 2013.

[7] M. Rohrbach, S. Amin, M. Andriluka and B. Schiele. A Database for Fine Grained Activity Detection of Cooking Activities. CVPR 2012.

Acknowledgements

We graciously thank the authors of the previous code releases and video benchmark for making them publicly available.