cd dockerfile
docker build -t cuda1101 .
docker run --gpus all --name city3dqa -v <code_path>:/workspace --ipc=host -it cuda1101 /bin/bash

conda create -n city3dqa python=3.8

pip install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html

pip install -r requirements.txt

If there is error in code, please try to adjust the transformers version to 4.29.2 or safetensors version to 0.3.0 to fix it.

cd lib/pointnet2
python setup.py install

Please download the feature data and QA file of City-3DQA.

Please download it and unzip it to data folder.

The raw point cloud data is from UrbanBIS.

Please download the feature data of scene graph from this link. Please download it and unzip it to data folder.

• sg_cityu/
  • data/
    • qa/
      • sentence_mode/
      • urban_mode/
    • sg/
    • urbanbis/
      • urbanbis_data/
        • Lihu_Area1_aligned_bbox.npy
        • …
  • dockerfile/
  • models/
  • scene_graph/
    • Lihu_Area1/
      • edges.pth
      • …
  • Scripts

You can train Sg-CityU with the following code

bash train.sh