3DGraphLLM

In this work, we propose 3DGraphLLM, a method for constructing a learnable representation of a 3D scene graph, which serves as input for LLMs to perform 3D vision-language tasks.

News

[2025.6] We are pleased to inform you that our paper has been accepted for poster presentation at ICCV 2025! 🎉

[2024.12] We release 3DGraphLLM pre-training on GT instance segmentation scene graphs

[2024.12] We release 3DGraphLLM paper code

🔥 Semantic relations boost LLM performance on 3D Referred Object Grounding and Dense Scene Captioning tasks

	ScanRefer		Multi3dRefer		Scan2Cap		ScanQA		SQA3D
	Acc@0.25	Acc@0.5	F1@0.25	F1@0.5	CIDEr@0.5	B-4@0.5	CIDEr	B-4	EM
Chat-Scene	55.5	50.2	57.1	52.3	77.1	36.3	87.7	14.3	54.6
3DGraphLLM Vicuna-1.5	58.6	53.0	61.9	57.3	79.2	34.7	91.2	13.7	55.1
3DGraphLLM LLAMA3-8B	62.4	56.6	64.7	59.9	81.0	36.5	88.8	15.9	55.9

🔨 Preparation

• Prepare the environment:

  conda create -n 3dgraphllm python=3.9.17
  conda activate 3dgraphllm
  conda install pytorch==2.2.1 torchvision==0.17.1 torchaudio==2.2.1 pytorch-cuda=11.8 -c pytorch -c nvidia
  pip install -r requirements.txt
  

• If you don't have root permissions to install java (needed for pycocoeval scripts for metrics such as BLEU and CIDER), install it with conda:

conda install -c conda-forge openjdk

• Download LLM backbone:

  • We use LLAMA3-8B-Instruct in our experiments, which can be downloaded from Hugging Face.

  • Change the llama_model_path in config.py to the path of LLAMA3-8B-Instruct.

• Annotations and extracted features:

  Please follow the instructions in preprocess.

🤖 Training and Inference

• Pre-training on GT instance segmentation scene graphs.

  • Modify run_gt_pretrain.sh:

    train_tag="scanrefer#scan2cap#scanqa#sqa3d#multi3dref#nr3d_caption#obj_align"
    val_tag="scanrefer#scan2cap#scanqa#sqa3d#multi3dref"
    evaluate=False
    

    Explanation of "train_tag" and "val_tag"

    • Use # to seperate different datasets

    • Datasets:

      • scanrefer: ScanRefer Dataset
      • scan2cap: Scan2Cap Dataset
      • scanqa: ScanQA Dataset
      • sqa3d: SQA3D Dataset
      • multi3dref: Multi3dRefer Dataset
      • nr3d_caption: A captioning dataset originated from Nr3D.
      • obj_align: A dataset originated from ScanRefer to align the object identifiers with object tokens.

  • Run: bash scripts/run_gt_pretrain.sh

• Training

  • Modify run.sh:

    train_tag="scanrefer#scan2cap#scanqa#sqa3d#multi3dref#nr3d_caption#obj_align"
    val_tag="scanrefer#scan2cap#scanqa#sqa3d#multi3dref"
    evaluate=False
    pretrained_path="outputs/llama3-8b-gt-pretrain-2/ckpt_00_28927.pth"
    

  • Run: bash scripts/run.sh

• Inference

  • Modify run.sh:

    val_tag="scanrefer#scan2cap#scanqa#sqa3d#multi3dref"
    evaluate=True
    pretrained_path="/path/to/pretrained_model.pth"
    

  • Run: bash scripts/run.sh

🚀 Demo

• Run: bash demo/run_demo.sh. You will be prompted to ask different queries about Scene 435 of ScanNet.

📪 Contact

If you have any questions about the project, please open an issue in this repository or send an email to Tatiana Zemskova.

📑 Citation

If you find this work helpful, please consider citing our work as:

@misc{zemskova20243dgraphllm,
      title={3DGraphLLM: Combining Semantic Graphs and Large Language Models for 3D Scene Understanding}, 
      author={Tatiana Zemskova and Dmitry Yudin},
      year={2024},
      eprint={2412.18450},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2412.18450}, 
}

😊 Acknowledgement

Thanks to the open source of the following projects:

Chat-Scene