DoraemonGPT: Toward Understanding Dynamic Scenes with Large Language Models
(Exemplified as A Video Agent)

Zongxin Yang  Guikun Chen  Xiaodi Li  Wenguan Wang  Yi Yang✉ 

ReLER, CCAI, Zhejiang University

^✉Corresponding Author

ICML 2024 (arXiv Preprint)

Project Page

Overview. Given a video with a question/task, DoraemonGPT first extracts a Task-related Symbolic Memory, which has two types of memory for selection: space-dominant memory based on instances and time-dominant memory based on time frames/clips. The memory can be queried by sub-task tools, which are driven by LLMs with different prompts and generate symbolic language (i.e., SQL sentences) to do different reasoning. Also, other tools for querying external knowledge or utility tools are supported. For planning, DoraemonGPT employs the MCTS Planner to decompose the question into an action sequence by exploring multiple feasible N solutions, which can be further summarized into an informative answer.

Setup and Configuration 🛠️

---

• Python Version: Python 3.9 or newer installed on your system.
• API Keys: Obtain API keys from one or more of the following services:
  • OpenAI (for access to models like GPT-3.5)
  • Google Cloud (for access to models like OCR)
  • Google Search (for access to search online)
• Python Dependencies: Install all necessary Python libraries as specified in the requirements.txt file. You can install these dependencies using the following command:

  pip install -r requirements.txt

• Other Model Choice: You have the option to use models from Hugging Face's Transformers library. Ensure you have the necessary credentials to use and download the model.

Installation Steps

1. Clone the repository 📦:

   git clone https://github.com/z-x-yang/DoraemonGPT.git

2. Opt for a virtual environment 🧹 and install the dependencies 🧑‍🍳:

   pip install -r requirements.txt

3. Set up your API key 🗝️:

   • Fill in config/inference/inference.yaml with your keys:

     openai:
       GPT_API_KEY: ["put your openai key here", ...]
     
     google_cloud:
       CLOUD_VISION_API_KEY: [...]
       QUOTA_PROJECT_ID: [...]

4. Download the checkpoints and bulid related project🧩:

   Thanks for the authors of these open source projects below for providing valuable pre-training models with outstanding performance🤝. When utilizing these models, users must strictly adhere to the authors' licensing agreements and properly cite the sources in published works.

   • download the pretrained model for action recognition

     mkdir checkpoints  
     cd ./checkpoints
     
     #download the pretrained model for action recognition
     wget https://pjlab-gvm-data.oss-cn-shanghai.aliyuncs.com/uniformerv2/k400/k400_k710_uniformerv2_b16_8x224.pyth

   • download the pretrained model for yolo-tracking

     #download the pretrained model for object detection and tracking
     wget https://objects.githubusercontent.com/github-production-release-asset-2e65be/521807533/0c7608ab-094c-4c63-8c0c-3e7623db6114?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=releaseassetproduction%2F20240612%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20240612T083947Z&X-Amz-Expires=300&X-Amz-Signature=7b6688c64e3d3f1eb54a0eca30ca99e140bed9f886d4c8a084bec389046ecda8&X-Amz-SignedHeaders=host&actor_id=0&key_id=0&repo_id=521807533&response-content-disposition=attachment%3B%20filename%3Dyolov8n-seg.pt&response-content-type=application%2Foctet-stream
     wget https://objects.githubusercontent.com/github-production-release-asset-2e65be/521807533/67360104-677c-457e-95a6-856f07ba3f2e?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=releaseassetproduction%2F20240612%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20240612T083803Z&X-Amz-Expires=300&X-Amz-Signature=8bd5d0f9ef518ee1a84783203b2d0a6c285a703dace053ae30596c68f2428599&X-Amz-SignedHeaders=host&actor_id=0&key_id=0&repo_id=521807533&response-content-disposition=attachment%3B%20filename%3Dyolov8n.pt&response-content-type=application%2Foctet-stream

   • download the pretrained model for dense captioning

     mkdir ./blip
     cd ./blip
     # dowlond the chekpoints from below 
     [[Hugging Face](https://huggingface.co/Salesforce/blip-image-captioning-large/tree/main)]
     cd ..

   • download the pretrained model for inpainting

     #download the pretrained model for inpainting
     mkdir ./E2FGVI
     cd ./E2FGVI
     # dowlond the chekpoints from below 
     [[Google Drive](https://drive.google.com/file/d/1tNJMTJ2gmWdIXJoHVi5-H504uImUiJW9/view?usp=sharing)] 
     [[Baidu Disk](https://pan.baidu.com/s/1qXAErbilY_n_Fh9KB8UF7w?pwd=lsjw)]
     cd ..

   • download the pretrained model for rvos

     #download the pretrained model for rvos
     mkdir AOT 
     cd ./AOT
     # dowlond the chekpoints from below 
     [[Google Drive](https://drive.google.com/file/d/1QoChMkTVxdYZ_eBlZhK2acq9KMQZccPJ/view)]
     cd ..
     
     mkdir GroundedSAM
     cd ./GroundedSAM
     wget https://objects.githubusercontent.com/github-production-release-asset-2e65be/611591640/c4c55fde-97e5-47d9-a2c5-b169832a2fa9?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=releaseassetproduction%2F20240623%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20240623T053405Z&X-Amz-Expires=300&X-Amz-Signature=369fd1d480eb018f7b3a31e960835ae77ae5bb9b1d0dcc5415751811daf4e325&X-Amz-SignedHeaders=host&actor_id=97865789&key_id=0&repo_id=611591640&response-content-disposition=attachment%3B%20filename%3Dgroundingdino_swinb_cogcoor.pth&response-content-type=application%2Foctet-stream
     # dowlond the chekpoints from below 
     [[Github](https://github.com/ChaoningZhang/MobileSAM/blob/master/weights/mobile_sam.pt)]
     cd ../..

---

QuickStart 🚀

• Fill in config/demo.yaml with your video and question:

  # run
  cd ..                  
  python ./scripts/demo_test.py --config config/demo_2.yaml  

---

News and Todo🗓️

• Release Code for Demo
• Release Code for Benchmarks
• Release Code with Video-LLaVA

---

Overview 📜

Thanks to the authors of these open source projects for providing excellent projects.

Memory Construction

• Time-dominant Memory ⏱️
  • Audio content
    • Whisper
      • https://github.com/openai/whisper
  • Captioning
    • BLIP/ BLIP2/ InstructBLIP
      • https://huggingface.co/Salesforce/blip-image-captioning-large
      • https://huggingface.co/Salesforce/blip2-opt-2.7b
      • https://huggingface.co/docs/transformers/model_doc/instructblip
  • Optical content
    • Google cloud
      • https://cloud.google.com/vision/docs/ocr?hl=zh-cn#vision_text_detection-python
• Space-dominant Memory 🌐
  • Category & Trajectory & Segmentation
    • BOXMOT(yolo-tracking)+yolov8-seg
      • mikel-brostrom/yolo_tracking: A collection of SOTA real-time, multi-object tracking algorithms for object detectors (github.com)
    • GroundedSAM+Deaot
      • https://github.com/IDEA-Research/Grounded-Segment-Anything
      • https://github.com/yoxu515/aot-benchmark
  • Appearance
    • BLIP/ BLIP2/ InstructBLIP
      • https://huggingface.co/Salesforce/blip-image-captioning-large
      • https://huggingface.co/Salesforce/blip2-opt-2.7b
      • https://huggingface.co/docs/transformers/model_doc/instructblip
  • Action Recognition
    • InternVideo
      • https://github.com/OpenGVLab/InternVideo

Tool Usage

• GoogleSearch
  • langchain - google
    • Google Search | 🦜️🔗 Langchain
• VideoInpainting
  • E2FGVI
    • MCG-NKU/E2FGVI: Official code for "Towards An End-to-End Framework for Flow-Guided Video Inpainting" (CVPR2022) (github.com)

---

Citations

Please consider citing the related paper(s) in your publications if it helps your research.

@inproceedings{yang2024doraemongpt,
  title={Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent)},
  author={Yang, Zongxin and Chen, Guikun and Li, Xiaodi and Wang, Wenguan and Yang, Yi},
  booktitle={Forty-first International Conference on Machine Learning}
}

License 🏷️

This project is all yours under the MIT License.