ArrangementNet

Official implementation of the paper ArrangementNet: Learning Scene Arrangements for Vectorized Indoor Scene Modeling. (SIGGRAPH 2023)

We provide three datasets at data/, which are consistent with the paper. cyberverse includes 54 large-scale scenes, floorsp is obtained from https://github.com/woodfrog/floor-sp, structured3d is obtained from https://github.com/bertjiazheng/Structured3D. We have already generated the arrangement graph based on three datasets and saved at data/*/arrangement_graph, which can be directly used as input to the network. We provide the evaluation groundtruth at data/*/evaluation_groundtruth.

1. Environment Setup

This repo was developed and tested with Python3.7, and you should install the following version of dgl and torch.

dgl==0.6.1

torch==1.8.0

2. complie gco

We rectify the network prediction by graphcut, so you should compile gco to get libgraphcut.so at src/gco/build.

cd src/gco
mkdi build && cd build
cmake ..
make -j8

3. train

python src/train.py --dataset cyberverse
python src/train.py --dataset floorsp
python src/train.py --dataset structured3d

4. inference

You can replicate the results in the paper using the pretrained model under checkpoint/.

python src/train.py --eval 1 --dataset cyberverse --resume checkpoint/cyberverse.ckpt
python src/train.py --eval 1 --dataset floorsp --resume checkpoint/floorsp.ckpt
python src/train.py --eval 1 --dataset structured3d --resume checkpoint/structured3d.ckpt

5. evaluate

The inference results will be saved at eval/, then you can evaluate the results using the following command to get the precision and recall of Room/Corner/Angle.

python evaluation/evaluations.py --predict_result eval/cyberverse --dataset cyberverse
python evaluation/evaluations.py --predict_result eval/floorsp --dataset floorsp
python evaluation/evaluations.py --predict_result eval/structured3d --dataset structured3d

For example, you will get evaluation result of cyberverse dataset like this

avg Room precision and recall 0.8804061910311912 0.8134793800970271

avg Corner precision and recall 0.8220827182273324 0.8274320067970539

avg Angle precision and recall 0.8004808248122176 0.8051524746935288