README.md

TGN

Tensorflow Implementation of the EMNLP-2018 paper Temporally Grounding Natural Sentence in Video by Jingyuan Chen et al.

Requirements

pip install -r requirements.txt

Data Preparation

1. Download Glove word embedding data.

cd download/
sh download_glove.sh

2. Download dataset features.

TACoS: BaiduDrive, GoogleDrive

Charades-STA: BaiduDrive, GoogleDrive

ActivityNet-Captions: BaiduDrive, GoogleDrive

Put the feature hdf5 file in the corresponding directory ./datasets/{DATASET}/features/

We decode TACoS/Charades videos using fps=16 and extract C3D (fc6) features for each non-overlap 16-frame snippet. Therefore, each feature corresponds to 1-second snippet. For ActivityNet, each feature corresponds to 2-second snippet. To extract C3D fc6 features, I mainly refer to this code.

3. Download trained models.

Download and put the checkpoints in corresponding ./checkpoints/{DATASET}/ .

BaiduDrive, GoogleDrive

4. Data Preprocessing (Optional)

cd datasets/tacos/
sh prepare_data.sh

Then copy the generated data in ./data/save/ .

Use correspondig scripts for preparing data for other datasets.

You may skip this procedure as the prepared data is already saved in ./datasets/{DATASET}/data/save/ .

Testing and Evaluation

sh scripts/test_tacos.sh
sh scripts/eval_tacos.sh

Use corresponding scripts for testing or evaluating for other datasets.

The predicted results are also provided in ./results/{DATASET}/ .

Training

sh scripts/train_tacos.sh

Use corresponding scripts for training for other datasets.