This project hosts the tensorflow implementation for our ECCV 2018 paper, A Joint Sequence Fusion Model for Video Question Answering and Retrieval}.

If you use this code or dataset as part of any published research, please refer the following paper.

@inproceedings{
  author    = {Youngjae Yu and Jongseok Kim and Gunhee Kim},
  title     = "{A Joint Sequence Fusion Model for Video Question Answering and Retrieval}"
  booktitle = {ECCV},
  year      = 2018
}

pip install -r requirements.txt

git submodule update --init --recursive
add2virtualenv .

• Video Feature

  1. Download LSMDC data.

  2. Extract rgb features using pool5 layer of the pretrained ResNet-152 model.

  3. Extract audio features using VGGish.

  4. Concat rgb and video features and save it into hdf5 file, and save it in 'dataset/LSMDC/LSMDC16_features/RESNET_pool5wav.hdf5'.

• Dataset

  • We processed raw data frames file in LSMDC17 and MSR-VTT dataset
  • Download dataframe files
  • Save these files in "dataset/LSMDC/DataFrame"

• Vocabulary

  • We make word embedding matrix using GloVe Vector.
  • Download vocabulary files
  • Save these files in "dataset/LSMDC/Vocabulary"

Modify configuartion.py to suit your environment.

• train_tag can be 'MC', 'FIB'

Run train.py.

python train.py --tag="tag"

You can download the models and features in gDrive Link Modify 'configuration.py' to load the checkpoints (self.load_from_ckpt = 'path/to/checkpoint/')

[RET] R@1: 93, R@5: 247, R@10: 348, medr : 29
[FIB] Accuracy: 45.1

You can get slightly lower or higher performance from these scores.