Awesome

News

• [Sep 21 2024]: 🎉🎉🎉🎉🎉 We are excited to release LHRS-Bot-Nova, a powerful new model with an upgraded base model, enriched training data, and an improved training recipe. Please check out the nova branch here and follow the instructions to start using the model. The model checkpoint is available on huggingface.
• [Jul 15 2024]: We updated our paper at arxiv.
• [Jul 12 2024]: We post the missing part in our paper (some observations, considerations, and lessons) in Zhihu (In Chinese, please contact us if you need English version).
• [Jul 09 2024]: We have released our evaluation benchmark LHRS-Bench.
• [Jul 02 2024]: Our paper has been accepted by ECCV 2024! We have open-sourced our training script and training data. Please follow the training instruction belown and data preparation.
• [Feb 21 2024]: We have updated our evaluation code. Any advice are welcom!
• [Feb 7 2024]: Model weights are now available on both Google Drive and Baidu Disk.
• [Feb 6 2024]: Our paper now is available at arxiv.
• [Feb 2 2024]: We are excited to announce the release of our code and model checkpoint! Our dataset and training recipe will be update soon!

Introduction

We are excited to introduce LHRS-Bot, a multimodal large language model (MLLM) that leverages globally available volunteer geographic information (VGI) and remote sensing images (RS). LHRS-Bot demonstrates a deep understanding of RS imagery and possesses the capability for sophisticated reasoning within the RS domain. In this repository, we will release our code, training framework, model weights, and dataset!

Preparation

Installation

1. Clone this repository.

   git clone git@github.com:NJU-LHRS/LHRS-Bot.git
   cd LHRS-Bot
   

2. Create a new virtual environment

   conda create -n lhrs python=3.10
   conda activate lhrs
   

3. Install dependences and our package

   pip install -e .
   

Checkpoints

• LLaMA2-7B-Chat

  • Automaticaly download:

    Our framework is designed to automatically download the checkpoint when you initiate training or run a demo. However, there are a few preparatory steps you need to complete:

    1. Request the LLaMA2-7B-Chat models from Hugging Face.

    2. After your request been processed, login to huggingface using your personal access tokens:

       huggingface-cli login
       (Then paste your access token and press Enter)
       

    3. Done!

  • Mannually download:

    • Download all the files from HuggingFace.

    • Change the following line of each file to your downloaded directory:

      • /Config/multi_modal_stage{1, 2, 3}.yaml

        ...
        text:
        	...
          path: ""  # TODO: Direct to your directory
        ...
        

      • /Config/multi_modal_eval.yaml

        ...
        text:
        	...
          path: ""  # TODO: Direct to your directory
        ...
        

• LHRS-Bot Checkpoints:

  Staeg1	Stage2	Stage3
  Baidu Disk, Google Drive	Baidu Disk, Google Drive	Baidu Disk, Google Drive

  • ⚠️ Ensure that the TextLoRA folder is located in the same directory as FINAL.pt. The name TextLoRA should remain unchanged. Our framework will automatically detect the version perceiver checkpoint and, if possible, load and merge the LoRA module.

  • Development Checkpoint:

    We will continually update our model with advanced techniques. If you're interested, feel free to download it and have fun :)

    Development
    Baidu Disk, Google Drive

Training

• Prepare and reformat your data following the instruction from here.

• Stage1

  • Fill the OUTPUT_DIR and DATA_DIR of script1.
  • cd Script; bash train_stage1.sh

• Stage2

  • Fill the OUTPUT_DIR and DATA_DIR of script1
  • Fill the MODEL_PATH for loading the stage1' checkpoint
  • cd Script; bash train_stage2.sh

• Stage3 is same as Stage2 except for different folder and script (here).

Demo

• Online Web UI demo with gradio:

  python lhrs_webui.py \
       -c Config/multi_modal_eval.yaml \           # config file
       --checkpoint-path ${PathToCheckpoint}.pt \  # path to checkpoint end with .pt
       --server-port 8000 \                        # change if you need
       --server-name 127.0.0.1 \                   # change if you need
       --share                                     # if you want to share with other
  

• Command line demo:

  python cli_qa.py \
       -c Config/multi_modal_eval.yaml \                 # config file
       --model-path ${PathToCheckpoint}.pt \             # path to checkpoint end with .pt
       --image-file ${TheImagePathYouWantToChat} \       # path to image file (Only Single Image File is supported)
       --accelerator "gpu" \                             # change if you need ["mps", "cpu", "gpu"]
       --temperature 0.4 \
       --max-new-tokens 512
  

• Inference:

  • Classification

    python main_cls.py \
         -c Config/multi_modal_eval.yaml \                 # config file
         --model-path ${PathToCheckpoint}.pt \             # path to checkpoint end with .pt
         --data-path ${ImageFolder} \                      # path to classification image folder
         --accelerator "gpu" \                             # change if you need ["mps", "cpu", "gpu"]
         --workers 4 \
         --enabl-amp True \
         --output ${YourOutputDir}                         # Path to output (result, metric etc.)
         --batch-size 8 \
    

  • Visual Grounding

    python main_vg.py \
         -c Config/multi_modal_eval.yaml \                 # config file
         --model-path ${PathToCheckpoint}.pt \             # path to checkpoint end with .pt
         --data-path ${ImageFolder} \                      # path to image folder
         --accelerator "gpu" \                             # change if you need ["mps", "cpu", "gpu"]
         --workers 2 \
         --enabl-amp True \
         --output ${YourOutputDir}                         # Path to output (result, metric etc.)
         --batch-size 1 \                                  # It's better to use batchsize 1, since we find batch inference
         --data-target ${ParsedLabelJsonPath}              # is not stable.
    

  • Visual Question Answering

    python main_vqa.py \
         -c Config/multi_modal_eval.yaml \                 # config file
         --model-path ${PathToCheckpoint}.pt \             # path to checkpoint end with .pt
         --data-path ${Image} \                            # path to image folder
         --accelerator "gpu" \                             # change if you need ["mps", "cpu", "gpu"]
         --workers 2 \
         --enabl-amp True \
         --output ${YourOutputDir}                         # Path to output (result, metric etc.)
         --batch-size 1 \                                  # It's better to use batchsize 1, since we find batch inference
         --data-target ${ParsedLabelJsonPath}              # is not stable.
         --data-type "HR"                                  # choose from ["HR", "LR"]
    

Acknowledgement

• We gratitude to the following repositories for their wonderful works:
  • LLaVA
  • MiniGPT-4
  • LLaMA

Statement

• If you find our work is useful, please give us 🌟 in GitHub and consider cite our paper:

  @InProceedings{10.1007/978-3-031-72904-1_26,
  author="Muhtar, Dilxat and Li, Zhenshi and Gu, Feng and Zhang, Xueliang and Xiao, Pengfeng",
  title="LHRS-Bot: Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model",
  booktitle="Computer Vision -- ECCV 2024",
  year="2025",
  publisher="Springer Nature Switzerland",
  address="Cham",
  pages="440--457",
  isbn="978-3-031-72904-1"
  }
  
  @article{li2024lhrs,
  title={LHRS-Bot-Nova: Improved Multimodal Large Language Model for Remote Sensing Vision-Language Interpretation},
  author={Li, Zhenshi and Muhtar, Dilxat and Gu, Feng and Zhang, Xueliang and Xiao, Pengfeng and He, Guangjun and Zhu, Xiaoxiang},
  journal={arXiv preprint arXiv:2411.09301},
  year={2024}
  }
  

• Licence: Apache