Implementation of VectorFloorSeg: Two-Stream Graph Attention Network for Vectorized Roughcast Floorplan Segmentation

Environment

conda install --yes --file requirements.txt

Install pyg following the instruction from official site, we recommend pyg==2.0.4

Data preparation

Download our processed data: here

Pretrained backbone downloaded

mkdir models
cd models

Download ResNet-101 from pytorch official site here, rename to resnet101-torch.pth and move to models.

Code preparation

Replace the graphgym and torch_geometric in pyg with corresponding dir in our repository

Train

python graphgym/train.py --cfg graphgym/configs/CUBI.yaml --seed 0

Eval

python graphgym/eval.py --cfg graphgym/configs/CUBI.yaml --eval train.epoch_resume 1 \
                          train.ckpt_prefix best val.extra_infos True seed 0

Optional: Processing svg format dataset from CubiCasa-5k source data

Notice: before running the code, please change the data dir within the code into your souce data dir

Source data downloaded: here
Download CubiCasa-5k source code and configure the environment: here
Put Replace_with_Cubicasa into CubiCasa-5k code repo
Process source model.svg into roughcast svg format floorplans:

python Replace_with_CubiCasa/roughcast_data_generation.py
Render svg floorplan into image:

python Replace_with_CubiCasa/ImgRasterization.py
Render image annotation of floorplans:

python Replace_with_CubiCasa/svg_loader.py
Draw the wireframe of svg floorplan and turn the wireframe and image annotation into mmseg format:

python DataPreparation/ImageProcessing_CubiCasa.py
Process svg floorplan as pickle file:

python SvgProcessing_CubiCasa.py

DrZiji/VecFloorSeg