LLMVA-GEBC: Large Language Model with Video Adapter for Generic Event Boundary Captioning

Code for the LOVEU@CVPR2023 Workshop Generic Event Boundary Captioning (GEBC) Chanllenge. Our proposed method achieved a 76.14 score on the test set and won the $1^{st}$ place in the challenge. The technical report can be found here.

Introduction

We proposes an effective model LLMVA-GEBC (Large Language Model with Video Adapter for Generic Event Boundary Captioning): (1) We utilize a pretrained LLM for generating human-like captions with high quality. (2) To adapt the model to the GEBC task, we take the video Q-former as an adapter and train it with the frozen visual feature extractors and LLM.

Enviroment Preparation

First, you should create a conda environment:

conda env create -f environment.yml
conda activate llmvagebc

Prerequisite Checkpoints

Before using the repository, make sure you have obtained the following checkpoints:

• pretrain-vicuna13b
• facebook/opt-13b

Remember to change the path of checkpoints ckpt in the config file.

Data

Download the Kinetic-GEBC dataset from https://sites.google.com/view/loveucvpr23/track2.

For primary visual feature: Using BLIP-2 to extract primary visual features. We use feature_extraction.py to do so. Remember to change the video_dir and save_dir in train_configs/blip2_feature_extract.yaml before you run:

python feature_extraction.py --cfg-path train_configs/blip2_feature_extract.yaml

For other visual features: CLIP to extract frame-level features and Omnivore to extract clip-level features. We use this pipeline to extract features.

Then, put the extracted features under these three folders:

data/features/eva_vit_g_q_former_tokens_12
data/features/clip_fps_15_stride_1_rename,
data/features/omnivore_fps_15_len_16_stride_1_rename

You can also directly download the official provided features here. But, remember to change the q_former_feature_folder, other_feat_total_size, other_feature_names and other_feature_folders in the config file.

Using VinVL to extract region-level features. The region feature of a video is saved to multiple .npy files, where each single file contains the region feature of a sampled frame. Merge the feature file paths into video_to_frame_index.json in the following format:

{
    "video_id": [
        "frame_1_feat.npy",
        "frame_2_feat.npy",
        ...     
    ],
    ...
}

Then put this file under data/features/.

Training and Validation

Firstly, set the configs in train_configs/${NAME_OF_YOUR_CONFIG_FILE}.yaml. Then run the script

CUDA_VISIBLE_DEVICES=${YOUR_GPU_ID} python train.py \
    --cfg-path train_configs/${NAME_OF_YOUR_CONFIG_FILE}.yaml

The results can be found in video_llama/output/.

Acknowledgement

We are grateful for the following awesome projects our LLMVA-GEBC arising from:

• Context-GEBC
• BLIP-2
• Video-LLaMA
• OPT
• Kinetic-GEBC

Citation

If you find our code useful, please cite the repo as follows:

@article{tang2023llmva,
  title={LLMVA-GEBC: Large Language Model with Video Adapter for Generic Event Boundary Captioning},
  author={Tang, Yunlong and Zhang, Jinrui and Wang, Xiangchen and Wang, Teng and Zheng, Feng},
  journal={arXiv preprint arXiv:2306.10354},
  year={2023}
}