This is the official implementation of the paper Semantic Fusion Augmentation and Semantic Boundary Detection: A Novel Approach to Multi-Target Video Moment Retrieval (WACV 2024).

• Install python packages.

  $ pip install -r requirements.txt

• (Optional) Boost mAP calculation.

  $ python setup.py install

• The datasets can be downloaded from our OneDrive.

• The folder structure should be like this:

  ./data
  ├── ActivityNet
  │   ├── C3D
  │   │   └── activitynet_v1-3_c3d.hdf5
  │   ├── I3D
  │   │   ├── v_00Dk03Jr70M.npy
  │   │   ├── v_00KMCm2oGhk.npy
  │   │   ├── ...
  |   |   └── ...
  │   ├── multi_test.json
  │   ├── test.json
  │   ├── train.json
  │   └── val.json
  ├── CharadesSTA
  │   ├── VGG
  │   │   └── vgg_rgb_features.hdf5
  │   ├── C3D
  │   │   └── Charades_C3D.hdf5
  │   ├── I3D
  │   │   ├── 001YG.npy
  │   │   ├── 003WS.npy
  │   │   ├── ...
  |   |   └── ...
  │   ├── multi_test.json
  │   ├── test.json
  │   └── train.json
  └── QVHighlights
      ├── features
      │   ├── clip_features
      │   ├── clip_text_features
      │   └── slowfast_features
      ├── test.json
      ├── train.json
      └── val.json
  

All our training configuration files are in the ./configs folder. The training command is as follows:

• Single GPU training.

  $ python main.py --config path/to/config.json --logdir path/to/log/dir

  For example, to train the model on CharadesSTA dataset with VGG backbone:

  $ python main.py --config ./configs/charades-VGG.json --logdir ./logs/charades-VGG-log

• Multi-GPU training. For example, to train the model on CharadesSTA dataset with VGG backbone on 4 GPUs:

  $ CUDA_VISIBLE_DEVICES=0,1,2,3 python main.py --config ./configs/charades-VGG.json --logdir ./logs/charades-VGG-log

The pretrained models used in our paper can be downloaded from OneDrive.

• It is recommended to put the pretrained models in the ./logs folder.

  ./logs
  ├── activity-C3D-log
  │   ├── ...
  ├── activity-I3D-log
  │   ├── ...
  ├── charades-C3D-log
  │   ├── ...
  ├── charades-I3D-log
  │   ├── ...
  ├── charades-VGG-log
  │   ├── best.pth
  │   └── config.json
  └── qv-log
      ├── ...
  

• Reproduce the results in our paper. Take the CharadesSTA dataset as an example:

  $ python main.py --test_only --config ./logs/charades-VGG-log/config.json  --logdir ./logs/charades-VGG-log

If you find this code useful for your research, please cite our paper:

@InProceedings{Huang_2024_WACV,
    author    = {Cheng Huang, Yi-Lun Wu, Hong-Han Shuai, Ching-Chun Huang},
    title     = {Semantic Fusion Augmentation and Semantic Boundary Detection: A Novel Approach to Multi-Target Video Moment Retrieval},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2024}
}