BuildingNet

This is the implementation of the BuildingNet architecture described in this paper:

Paper:

BuildingNet: Learning to Label 3D Buildings

Arxiv version:

(https://arxiv.org/abs/2110.04955)

Project Page:

https://buildingnet.org/

Requirements:

This project was built using cuda10.1 and python3.8
For other requirements, look into requirements.txt. The conda environment is in 'buildingnet.yml'

Model features:

The model features are combinations of a pretrained network model features and building prior information features.
In this paper we have used minkowskiNet to train for the pretrained features.
Minkowski CNN

Run the model:

1. After downloading the dataset (fill in the form on our official project page to get access) place the contents of model_data/GNN under the data folder in the project

2. To run this model, execute command in run.txt
   

python3 train.py --datadir="data" --epoch 200 --outputdir 'Output' --ckpt_dir 'checkpoint' --normalization 'GN' --modeltype 'Edge' --edgetype 'all' --lr 1e-4 --nodetype 'node+minkownormal' --pretrainedtype 'fc3_avg' --IOU_checkpoint=5

This gives shape and part IOU every 5 epochs