【CVPR'2023 Highlight🔥】Video-Text as Game Players: Hierarchical Banzhaf Interaction for Cross-Modal Representation Learning

The implementation of CVPR 2023 Highlight (Top 10%) paper Video-Text as Game Players: Hierarchical Banzhaf Interaction for Cross-Modal Representation Learning.

In this paper, we creatively model video-text as game players with multivariate cooperative game theory to wisely handle the uncertainty during fine-grained semantic interaction with diverse granularity, flexible combination, and vague intensity.

📌 Citation

If you find this paper useful, please consider staring 🌟 this repo and citing 📑 our paper:

@inproceedings{jin2023video,
  title={Video-text as game players: Hierarchical banzhaf interaction for cross-modal representation learning},
  author={Jin, Peng and Huang, Jinfa and Xiong, Pengfei and Tian, Shangxuan and Liu, Chang and Ji, Xiangyang and Yuan, Li and Chen, Jie},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={2472--2482},
  year={2023}
}

💡 I also have other text-video retrieval projects that may interest you ✨.

DiffusionRet: Generative Text-Video Retrieval with Diffusion Model
Accepted by ICCV 2023 | [DiffusionRet Code]
Peng Jin, Hao Li, Zesen Cheng, Kehan Li, Xiangyang Ji, Chang Liu, Li Yuan, Jie Chen

Expectation-Maximization Contrastive Learning for Compact Video-and-Language Representations
Accepted by NeurIPS 2022 | [EMCL Code]
Peng Jin, Jinfa Huang, Fenglin Liu, Xian Wu, Shen Ge, Guoli Song, David Clifton, Jie Chen

Text-Video Retrieval with Disentangled Conceptualization and Set-to-Set Alignment
Accepted by IJCAI 2023 | [DiCoSA Code]
Peng Jin, Hao Li, Zesen Cheng, Jinfa Huang, Zhennan Wang, Li Yuan, Chang Liu, Jie Chen

📣 Updates

• [2023/10/15]: We release our pre-trained estimator weights. If you want to apply a to other tasks, you can initialize a new estimator with the weights we provide. If you want better performance, you can train the estimator with a smaller learning rate and more epochs.
• [2023/10/11]: We release code for Banzhaf Interaction estimator. Recommended running parameters will be provided shortly, and we will also release our pre-trained estimator weights.
• [2023/10/08]: I am working on the code for Banzhaf Interaction estimator, which is expected to be released soon.
• [2023/06/28]: Release code for reimplementing the experiments in the paper.
• [2023/03/28]: Our HBI has been selected as a Highlight paper at CVPR 2023! (Top 2.5% of 9155 submissions).
• [2023/02/28]: We will release the code asap. (I am busy with other DDLs. After that, I will open the source code as soon as possible. Please understand.)

⚡ Demo

demo_github.mp4

😍 Visualization

Example 1

More examples

Example 2

Example 3

Example 4

Example 5

Example 6

Example 7

🚀 Quick Start

Setup

Setup code environment

conda create -n HBI python=3.9
conda activate HBI
pip install -r requirements.txt
pip install torch==1.8.1+cu102 torchvision==0.9.1+cu102 -f https://download.pytorch.org/whl/torch_stable.html

Download CLIP Model

cd HBI/models
wget https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt
# wget https://openaipublic.azureedge.net/clip/models/5806e77cd80f8b59890b7e101eabd078d9fb84e6937f9e85e4ecb61988df416f/ViT-B-16.pt
# wget https://openaipublic.azureedge.net/clip/models/b8cca3fd41ae0c99ba7e8951adf17d267cdb84cd88be6f7c2e0eca1737a03836/ViT-L-14.pt

Download Datasets

Datasets	Google Cloud	Baidu Yun	Peking University Yun
MSR-VTT	Download	Download	Download
MSVD	Download	Download	Download
ActivityNet	TODO	Download	Download
DiDeMo	TODO	Download	Download

Train the Banzhaf Interaction Estimator

Train the estimator according to the label generated by the BanzhafInteraction in HBI/models/banzhaf.py.

The training code is provided in banzhaf_estimator.py. We provide our trained weights, and if you want to apply a to other tasks, you can initialize a new estimator with the weights we provide.

We have tested the performance of Estimator_1e-2_epoch6 with R@1 of 48.2 (log) on the MSR-VTT dataset. If you want better performance, you can train the estimator with a smaller learning rate and more epochs.

Models	Google Cloud	Baidu Yun	Peking University Yun	log
Estimator_1e-2_epoch1	Download	Download	Download	log
Estimator_1e-2_epoch2	Download	Download	Download	log
Estimator_1e-2_epoch3	Download	Download	Download	log
Estimator_1e-2_epoch4	Download	Download	Download	log
Estimator_1e-2_epoch5	Download	Download	Download	log
Estimator_1e-2_epoch6	Download	Download	Download	log

CUDA_VISIBLE_DEVICES=0,1,2,3 \
python -m torch.distributed.launch \
--master_port 2502 \
--nproc_per_node=4 \
banzhaf_estimator.py \
--do_train 1 \
--workers 8 \
--n_display 1 \
--epochs 10 \
--lr 1e-2 \
--coef_lr 1e-3 \
--batch_size 128 \
--batch_size_val 128 \
--anno_path data/MSR-VTT/anns \
--video_path ${DATA_PATH}/MSRVTT_Videos \
--datatype msrvtt \
--max_words 24 \
--max_frames 12 \
--video_framerate 1 \
--output_dir ${OUTPUT_PATH} 

Text-video Retrieval

Checkpoint	Google Cloud	Baidu Yun	Peking University Yun
MSR-VTT	Download	Download	Download
ActivityNet	Download	Download	Download

Eval on MSR-VTT

CUDA_VISIBLE_DEVICES=0,1 \
python -m torch.distributed.launch \
--master_port 2502 \
--nproc_per_node=2 \
main_retrieval.py \
--do_eval 1 \
--workers 8 \
--n_display 50 \
--batch_size_val 128 \
--anno_path data/MSR-VTT/anns \
--video_path ${DATA_PATH}/MSRVTT_Videos \
--datatype msrvtt \
--max_words 24 \
--max_frames 12 \
--video_framerate 1 \
--init_model ${CHECKPOINT_PATH} \
--output_dir ${OUTPUT_PATH} 

Train on MSR-VTT

CUDA_VISIBLE_DEVICES=0,1 \
python -m torch.distributed.launch \
--master_port 2502 \
--nproc_per_node=2 \
main_retrieval.py \
--do_train 1 \
--workers 8 \
--n_display 50 \
--epochs 5 \
--lr 1e-4 \
--coef_lr 1e-3 \
--batch_size 128 \
--batch_size_val 128 \
--anno_path data/MSR-VTT/anns \
--video_path ${DATA_PATH}/MSRVTT_Videos \
--datatype msrvtt \
--max_words 24 \
--max_frames 12 \
--video_framerate 1 \
--estimator ${ESTIMATOR_PATH} \
--output_dir ${OUTPUT_PATH} \
--kl 2 \
--skl 1

Eval on ActivityNet Captions

CUDA_VISIBLE_DEVICES=0,1 \
python -m torch.distributed.launch \
--master_port 2502 \
--nproc_per_node=2 \
main_retrieval.py \
--do_eval 1 \
--workers 8 \
--n_display 50 \
--batch_size_val 128 \
--anno_path ${DATA_PATH}/ActivityNet \
--video_path ${DATA_PATH}/ActivityNet/Activity_Videos \
--datatype activity \
--max_words 64 \
--max_frames 64 \
--video_framerate 1 \
--init_model ${CHECKPOINT_PATH} \
--output_dir ${OUTPUT_PATH} 

Train on ActivityNet Captions

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
python -m torch.distributed.launch \
--master_port 2502 \
--nproc_per_node=8 \
main_retrieval.py \
--do_train 1 \
--workers 8 \
--n_display 10 \
--epochs 10 \
--lr 1e-4 \
--coef_lr 1e-3 \
--batch_size 128 \
--batch_size_val 128 \
--anno_path ${DATA_PATH}/ActivityNet \
--video_path ${DATA_PATH}/ActivityNet/Activity_Videos \
--datatype activity \
--max_words 64 \
--max_frames 64 \
--video_framerate 1 \
--estimator ${ESTIMATOR_PATH} \
--output_dir ${OUTPUT_PATH} \
--kl 2 \
--skl 1

Video-question Answering

Checkpoint	Google Cloud	Baidu Yun	Peking University Yun
MSR-VTT-QA	Download	Download	Download

Eval on MSR-VTT-QA

CUDA_VISIBLE_DEVICES=0,1 \
python -m torch.distributed.launch \
--master_port 2502 \
--nproc_per_node=2 \
main_vqa.py \
--do_eval \ 
--num_thread_reader=8 \
--train_csv data/MSR-VTT/qa/train.jsonl \
--val_csv data/MSR-VTT/qa/test.jsonl \
--data_path data/MSR-VTT/qa/train_ans2label.json \
--features_path ${DATA_PATH}/MSRVTT_Videos \
--max_words 32 \
--max_frames 12 \
--batch_size_val 16 \
--datatype msrvtt \
--expand_msrvtt_sentences  \
--feature_framerate 1 \
--freeze_layer_num 0  \
--slice_framepos 2 \
--loose_type \
--linear_patch 2d \
--init_model ${CHECKPOINT_PATH} \
--output_dir ${OUTPUT_PATH}

Train on MSR-VTT-QA

CUDA_VISIBLE_DEVICES=0,1 \
python -m torch.distributed.launch \
--master_port 2502 \
--nproc_per_node=2 \
main_vqa.py \
--do_train \ 
--num_thread_reader=8 \
--epochs=5 \
--batch_size=32 \
--n_display=50 \
--train_csv data/MSR-VTT/qa/train.jsonl \
--val_csv data/MSR-VTT/qa/test.jsonl \
--data_path data/MSR-VTT/qa/train_ans2label.json \
--features_path ${DATA_PATH}/MSRVTT_Videos \
--lr 1e-4 \
--max_words 32 \
--max_frames 12 \
--batch_size_val 16 \
--datatype msrvtt \
--expand_msrvtt_sentences  \
--feature_framerate 1 \
--coef_lr 1e-3 \
--freeze_layer_num 0  \
--slice_framepos 2 \
--loose_type \
--linear_patch 2d \
--estimator ${ESTIMATOR_PATH} \
--output_dir ${OUTPUT_PATH} \
--kl 2 \
--skl 1

🎗️ Acknowledgments

Our code is based on EMCL, CLIP, CLIP4Clip and DRL. We sincerely appreciate for their contributions.