Action Knowledge for Video Captioning with Graph Neural Networks

Description

This is the implementation of our paper entitled "Action Knowledge for Video Captioning with Graph Neural Networks".

Our approach for video captioning introduces a new technique that leverages action as edge features within a graph neural network (GNN), with objects represented as nodes. By integrating object-action relationships into the GNN, our method enhances the visual representation and generates more precise captions. Furthermore, we enhance the performance by combining the proposed edge representation with a node representation based on grids. By overlapping the grids, the model captures more comprehensive information about the objects, leading to further improvements in performance.

It is demonstrated in the experiments of MSVD and MSR-VTT that our method improved video captioning quantitatively and qualitatively.

The illustration of our proposed action-graph model with overlapping grid is shown below:

Prepare the Environment

Install and create conda environment with the provided environment.yml file. This conda environment was tested with the NVIDIA A6000 and NVIDIA RTX 3090.

The details of each dependency can be found in the environment.yml file.

conda env create -f environment.yml
conda activate action_graph_env
pip install git+https://github.com/Maluuba/nlg-eval.git@master
pip install pycocoevalcap

Install torch following this page: https://pytorch.org/get-started/locally 
pip install opencv-python
pip install seaborn

Prepare the Dataset

Dataset Folder Structure

├── dataset
│   ├── MSVD
│   │   ├── raw # put the 1970 raw videos in here
│   │   ├── captions 
│   │   ├── raw-captions_mapped.pkl # mapping between video id with captions
│   │   ├── train_list_mapping.txt
│   │   ├── val_list_mapping.txt
│   │   ├── test_list_mapping.txt
│   ├── MSRVTT
│   │   ├── raw # put the 10000 raw videos in here
│   │   ├── msrvtt.csv # list of video id in msrvtt dataset
│   │   ├── MSRVTT_data.json # metadata of msrvtt dataset, which includes video url, video id, and caption

MSR-VTT

Raw videos can be downloaded from this link. We provided the captions in dataset/MSRVTT folder.

MSVD

Raw videos can be downloaded from this link. We provided the captions in dataset/MSVD folder.

Extract the Features

Feature Extractor Folder Structure

├── model
│   ├── i3d
├── modules # Copy the files from CLIP4Clip repository: https://github.com/ArrowLuo/CLIP4Clip/tree/master/modules
├── pretrained 
│   ├── [trained CLIP4Clip model].bin # Train your own PyTorch CLIP4Clip model
│   ├── rgb_imagenet.pt # Download I3D model: https://github.com/piergiaj/pytorch-i3d/blob/master/models/rgb_imagenet.pt
├── utility # Some helper functions to generate the features

Note

- Please make sure you have copied modules from CLIP4Clip https://github.com/ArrowLuo/CLIP4Clip/tree/master/modules into feature_extractor/modules.
- Please make sure you have downloaded rgb_imagenet.pt into feature_extractor/pretrained.
- Please change args in each notebook based on requirement e.g,. args.msvd = False for MSR-VTT and args.msvd = True for MSVD.

CLIP-based features

Steps:

1. Train CLIP4Clip based on https://github.com/ArrowLuo/CLIP4Clip and put the best model in the pretrained folder.
   • For MSR-VTT, we use 6513 clips for training, 497 clips for validation and 2990 clips for test when training the CLIP4Clip.
   • For MSVD, we use 1200 clips for training, 100 clips for validation and 670 clips for test when training the CLIP4Clip.
2. Extract the CLIP-based features by using clip4clip_theta_2_feature_extraction.ipynb.

Features of Grid-based Action Graph and Object-based Action Graph

Grid Based Action Graph

Steps:

1. Extract grid node by using grid_node_theta_1_feature_extractor.ipynb.

2. Extract spatial action graph by using grid_based_spatial_action_graph.ipynb.

3. Extract temporal action graph by using temporal_similarity_graph.ipynb.

4. Create the grid based action graph:

   • run action_spatio_temporal_graph_feature_extractor.ipynb then ,
   • run transform-graph-to-geometric.ipynb

Object Based Action Graph

Steps:

1. Extract object node by using object_node_theta_1_feature_extractor.ipynb.
2. Extract spatial action graph by using object_based_spatial_action_graph.ipynb.
3. Extract temporal action graph by using temporal_similarity_graph.ipynb.
4. Create the object based action graph:
   • Run action_spatio_temporal_graph_feature_extractor.ipynb then,
   • Run transform-graph-to-geometric.ipynb.

Training

1. Download a pretrained BERT model. This is used as word embedding and tokenizer for the captions.

mkdir modules/bert-model
cd modules/bert-model/
wget https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased-vocab.txt
mv bert-base-uncased-vocab.txt vocab.txt
wget https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased.tar.gz
tar -xvf bert-base-uncased.tar.gz
rm bert-base-uncased.tar.gz
cd ../../

2. Download a pretrained weight of UniVL. This is used to initialize our caption generator.

mkdir -p ./weight
wget -P ./weight https://github.com/microsoft/UniVL/releases/download/v0/univl.pretrained.bin

3. Prepare the CLIP-based features as mentioned in section [Extract the Features] -> [CLIP-based features].
4. Prepare the graph features, i.e., Grid-based Action Graph or Object-based Action Graph, as mentioned in section [Extract the Features] -> [Features of Grid-based Action Graph and Object-based Action Graph].
5. Open a train script (.sh) in folder scripts, and change following parameters based on the specs of your machine:
   • N_GPU = [Total GPU to use]
   • N_THREAD = [Total thread to use]
6. If needed, change also the following parameters according to the location of the data in your machine:
   • DATA_PATH = [MSR-VTT JSON file location] or [MSVD dataset location]
   • CKPT_ROOT = [Your desired folder for saving the models and results]
   • INIT_MODEL_PATH = [UniVL pretrained model location]
   • FEATURES_PATH = [CLIP-based features]
   • DATA_GEOMETRIC_PATH = [Generated Action Graph feature path (Grid-based action graph or Object-based action graph)]
7. Based on object detection model, change node_feat_dim according to the object feature dimension, e.g for YOLO the node_feat_dim is 1024.
8. Execute the following scripts to start the training process:

Train our proposed method

MSVD

cd scripts/
./msvd_train_GNN.sh 

MSRVTT

cd scripts/
./msrvtt_train_GNN.sh  

Evaluation

After the training is done, an evaluation process on the test set will be automatically executed using the best checkpoint among all epochs. However, if you want to evaluate a checkpoint at a specific epoch, you can use the provided training shell script by modifying the value of INIT_MODEL_PATH to the location of the desired checkpoint, and replacing the --do_train to --do_eval.

Our Results

The comparison with the existing methods and also the ablation study of our method can be found in our paper.

MSVD

Method	CLIP Model	BLEU@4	METEOR	ROUGE-L	CIDEr
Ours (Action + Object)	ViT-B/32	62.56	41.53	78.62	120.64
Ours (Action + Grid)	ViT-B/32	62.90	41.81	78.80	119.07
Ours (Action + Grid)	ViT-B/16	64.07	42.41	79.72	124.18

MSR-VTT

Method	CLIP Model	BLEU@4	METEOR	ROUGE-L	CIDEr
Ours (Action + Object)	ViT-B/32	48.31	31.35	65.34	60.00
Ours (Action + Grid)	ViT-B/32	49.10	31.57	65.52	61.27
Ours (Action + Grid)	ViT-B/16	51.02	32.19	66.55	63.02

Acknowledgements

Our code is developed based on https://github.com/microsoft/UniVL, which is also developed based on https://github.com/huggingface/transformers/tree/v0.4.0 and https://github.com/antoine77340/howto100m .

Citation

Please cite our paper in your publications if it helps your research as follows:

@article{Hendria2023,
      author   = {W. F. Hendria and V. Velda and B. H. H. Putra and F. Adzaka and C. Jeong},
      title    = {Action Knowledge for Video Captioning with Graph Neural Networks},
      journal  = {J. King Saud Univ.-Comput. Inf. Sci.},
      volume   = {35},
      number   = {4},
      pages    = {50-62},
      month    = apr,
      year     = {2023}"
}