DMC-Net

Citing

If you find DMC-Net useful, please consider citing:

@inproceedings{shou2019dmc,
	title={DMC-Net: Generating Discriminative Motion Cues for Fast Compressed Video Action Recognition},
	author={Shou, Zheng and Lin, Xudong and Kalantidis, Yannis and Sevilla-Lara, Laura and Rohrbach, Marcus and Chang, Shih-Fu and Yan, Zhicheng},
	booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
	year={2019}
}

Overview

./exp_my/ contains scripts for running experiments and our trained models and prediction results.

./code/ contains implementation for 3 major models respectively: 0. dmcnet indicates the version which does not include the adversarial loss during training and uses ResNet-18 for classifying DMC;

1. dmcnet_GAN indicates the version which includes the adversarial loss during training and uses ResNet-18 for classifying DMC;
2. dmcnet_I3D indicates the version which uses I3D for classifying DMC.

In the following, we present how to use dmcnet and dmcnet_GAN. Instructions for dmcnet_I3D can be found in ./code/dmcnet_I3D/.

DMC-Net with ResNet-18 classifier

Installation

We implement dmcnet and dmcnet_GAN using PyTorch based on CoViAR. Please refer to CoViAR for details of setup and installation (e.g. how to prepare input videos, setup mpeg-4 compressed video data loader, etc.). Specifically, the released models were trained using python 3.6, pytorch 0.31, cuda 9.0, MPEG-4 video of GOP 12 and macroblock size 16x16.

Optical flow extraction: we extract optical flow using TV-L1 algorithm implementation from dense_flow and store the flow images beforehand and then load flow images during training.

In both ./code/dmcnet/ and ./code/dmcnet_GAN/, please first link exp/ to true directory of './exp_my/' so that all data will be stored in the experimental folder.

Usage

As stated in the paper, we first train DMC-Net with the classification loss and flow reconstruction MSE loss but without the adversarial loss (using ./code/dmcnet/). Sample training script for HMDB-51 can be found at exp_my/hmdb51_gen_flow/split1/run.sh. Performing training and testing by exp/hmdb51_gen_flow/split1/run.sh;. The trained model would be exp/hmdb51_gen_flow/split1/_mv_model_best.pth.tar.

Explanations about some key options used in the run.sh script (detaied descriptions can be found in train_options.py):

0. data-root: specify the directory for storing mpeg-4 videos;

1. train-list and test-list: specify the training and testing videos lists. Some example lines in such list files (format follows CoViAR: directory class class_index):

   smile/Me_smiling_smile_h_nm_np1_fr_goo_0.avi smile 0

   clap/Alex_applauding_himself_clap_u_nm_np1_fr_med_0.avi clap 1

   climb/Chiara_Kletterwand_climb_f_cm_np1_ba_bad_0.avi climb 2

2. flow-root: specify the directory for storing ground truth optical flow images extracted by dense_flow. Sample directory:

   flow-root/brush_hair/April_09_brush_hair_u_nm_np1_ba_goo_0/flow_x_00001.jpg

   flow-root/brush_hair/April_09_brush_hair_u_nm_np1_ba_goo_0/flow_y_00001.jpg

   flow-root/brush_hair/April_09_brush_hair_u_nm_np1_ba_goo_0/flow_x_00002.jpg

   flow-root/brush_hair/April_09_brush_hair_u_nm_np1_ba_goo_1/flow_x_00001.jpg

   flow-root/climb_stairs/BIG_FISH_climb_stairs_f_nm_np1_fr_med_1/flow_x_00001.jpg

Then we use the above trained model as initialization to train with the adversarial loss included (using ./code/dmcnet_GAN/). Sample training script for HMDB-51 can be found at exp_my/hmdb51_gan/split1/run.sh. In order to reproduce the result on HMDB-51, simply run: bash exp/hmdb51_gan/split1/run.sh; bash ./exp/hmdb51_gan/split1/run_combine.sh 2>&1 | tee ./exp/hmdb51_gan/split1/acc.log The trained model would be exp/hmdb51_gan/split1/_mv_model_best.pth.tar and the prediction results would be stored in exp/hmdb51_gan/split1/mv_score_model_best.npz and ./exp/hmdb51_gan/split1/acc.log records the accuracy after fusing all modalities.

Our trained models

At AWS host (dl.fbaipublicfiles.com/dmc-net/models.zip), we provide our trained models and prediction results. The file directory of ./models/ follows similar structure as ./exp_my/. Please put the trained model and prediction result (for each dataset and split from ./models/) in the corresponding folder (for experiment in ./exp_my/).

Results

Accuracy (%)	HMDB-51	UCF-101
EMV-CNN	51.2 (split1)	86.4
DTMV-CNN	55.3	87.5
CoViAR	59.1	90.4
DMC-Net (ResNet-18)	62.8	90.9
DMC-Net (I3D)	71.8	92.3
DMC-Net (I3D) + I3D RGB	77.8	96.5

License

DMC-Net is MIT licensed, as found in the LICENSE file.