This repo holds the implementation code of the paper:
Optical Flow Guided Feature: A Fast and Robust Motion Representation for Video Action Recognition, Shuyang Sun, Zhanghui Kuang, Lu Sheng, Wanli Ouyang, Wei Zhang, CVPR 2018.
- OpenCV 2.4.12
- OpenMPI 1.8.5 (enable-thread-multiple when install)
- CUDA 7.5
- CUDNN 5
- Caffe Dependencies
You may refer to the project TSN to install these libs and prepare the data.
For training use, first modify the file make_train.sh with your own lib path filled in. Simply run sh make_train.sh, the script will automatically build the caffe for you.
For testing, you can simply run make pycaffe to make all stuff well prepared.
You need to make two folders before you launch your training. The one is logs under the root of this project, and the other is the model under the folder models/DATASET/METHOD/SPLIT/. For instance, if you want to train RGB_OFF on the dataset UCF101 split 1, then your model directory should be made under the path models/ucf101/RGB_OFF/1/.
The network structure for training is defined in train.prototxt, and the hyperparameters are defined in solver.prototxt. For detailed training strategies and observations not included in the paper, please refer to our training recipes.
You need to create another directory proto_splits under the same folder of model. Our test code use pycaffe to call the functions defined in C++, therefore, we need to write some temporary files for synchronization. Remember to clean the temporary files everytime you launch a new test. Run the script test.sh with your METHOD, MODEL_NAME, SPLIT and NUM_GPU specified.
The deploy_tpl.prototxt defines the network for reference. To transfer your network defined in train.prototxt into deploy_tpl.prototxt, you may need to copy all the layers except the data layer and layers after each fully connected layer. As there are dynamic parameters defined in the deploy_tpl.prototxt, e.g. $SOURCE $OVERSAMPLE_ID_PATH, the format of the deploy_tpl.prototxt is a little bit different to the normal prototxt file.
After testing, an aggregation operation is needed to fuse the scores from different sources. The script ensemble_test.sh may help you to aggregate results with manually searched weights (though inelegant:(......). You can find those weights settings among the comments of the script ensemble_test.sh.
Due to the unexpected server migration, our original models trained on all 3 splits of UCF101 and HMDB51 were all lost. Therefore, we re-train the models on the first split of UCF101:
| RGB | OFF(RGB) | RGB DIFF | OFF(RGB DIFF) | FLOW | OFF(FLOW) | Acc. (Acc. in Paper) |
|---|---|---|---|---|---|---|
| 90.5% (90.0%) | ||||||
| 93.2% (93.0%) | ||||||
| 95.3% (95.5%) |
| Model Name | Init Model | Reference Model |
|---|---|---|
| OFF(RGB) | Baidu Pan Google Drive |
Baidu Pan Google Drive |
| OFF(RGB DIFF) | Baidu Pan Google Drive |
Baidu Pan Google Drive |
| OFF(Flow) | Baidu Pan Google Drive |
Baidu Pan Google Drive |
If you find our research useful, please cite the paper:
@InProceedings{Sun_2018_CVPR,
author = {Sun, Shuyang and Kuang, Zhanghui and Sheng, Lu and Ouyang, Wanli and Zhang, Wei},
title = {Optical Flow Guided Feature: A Fast and Robust Motion Representation for Video Action Recognition},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2018}
}
The codes in this repo are published under the MIT License.
You can contact Mr.Shuyang Sun (Please do NOT call me with the title Prof., Dr., or any other kinds of weird prefixes. I'm still a master student....) by sending email to [email protected]