C3D: Generic Features for Video Analysis

Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, Manohar Paluri

C3D is a modified version of BVLC caffe [2] to support 3-Dimensional Convolutional Networks. C3D can be used to train, test, or fine-tune 3D ConvNets efficiently. We also provide our C3D pre-trained model which were trained on Sports-1M dataset [3] with necessary tools for extract video features.

Demos

Sport classification using C3D on Sports-1M dataset. Video frames are visualized with top 2 predictions.

Downloads

03/01/2017: New code and model is released (check out at git hub). Paper/techincal report will be availble soon.
Source code: github (we branched out from BVLC caffe on July 17, 2014 with gist_id b80fc86. Following normal caffe installation should work).
Pre-trained model: sport1m (alternative url1 or url2). For full reference of the model, please refer to [1]. For a quick reference, please refer to the table below where we report accuracy of this model C3D with fc6 feature and linear SVM.

Dataset	UCF101	ASLAN	UMD-Scene	YUPENN-Scene	Object
C3D + linear SVM	82.3	78.3 (86.5)	87.7	98.1	22.3

Usage

C3D User Guide.

Supports

This project is supported by:

Citation

If you find C3D helpful for your research, please cite the following papers:

D. Tran, L. Bourdev, R. Fergus, L. Torresani, and M. Paluri, Learning Spatiotemporal Features with 3D Convolutional Networks, ICCV 2015, PDF.
Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell, Caffe: Convolutional Architecture for Fast Feature Embedding, arXiv 2014.
A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and L. Fei-Fei, Large-scale Video Classification with Convolutional Neural Networks, CVPR 2014.

Need further helps? Email me: trandu -at- fb.com or post here.