Vectorizing World Buildings

Code and instructions for our paper: Vectorizing World Buildings: Planar Graph Reconstruction by Primitive Detection and Relationship Classification, ECCV 2020.

Data

We borrow images from SpaceNet [1] corpus, which are hosted as an Amazon Web Services (AWS) Public Dataset and annotated 2,001 buildings across Los Angeles, Las Vegas and Paris, click here to download our annotations and cropped RGB images.

[1] SpaceNet on Amazon Web Services (AWS). “Datasets.” The SpaceNet Catalog. Last modified April 30, 2018.

Running pretrained models

PS.: This code is still not clean nor optimized for good performance. Some modules were borrowed from Mask-RCNN, for more detailed documentation refer to https://github.com/facebookresearch/maskrcnn-benchmark.

Detecting corners primitives (PC) and corner-edge relationships (CE)

Change paths in refs.py and datasets/junction.py
Download pretrained model and move to output/{EXP_ID}/
Pretrained model https://www.dropbox.com/s/h72ux3w32o6t9au/pretrained_junctions.zip?dl=0
Run python3 main.py --exp 3 --json --test --checkepoch 15 --gpu 0
Predictions should appear in result/

Detecting edges primitives (PE)

Change paths in detect.py
Download pretrained model https://www.dropbox.com/s/b2dcqhb0de2xkua/pretrained_edges.zip?dl=0
Run python3 detect.py
Predictions will appear in ./output

Detecting regions primitive (PR)

Build maskrcnn
Download pretrained model https://www.dropbox.com/s/wclufzt3liq120y/pretrained_regions.zip?dl=0
Set paths in '/home/nelson/Workspace/outdoor_project_to_submit/region_detector/maskrcnn_benchmark/config/paths_catalog.py'
python3 ./tools/test_net.py --config-file '/home/nelson/Workspace/building_reconstruction/working_model/maskrcnn-boundary/configs/buildings_mask_rcnn_R_50_FPN_1x.yaml'

Detecting region-to-region relationships (RR)

Build maskrcnn
Download pretrained model https://www.dropbox.com/s/9ju3iwwexecz69j/pretrained_shared_edges.zip?dl=0
Set paths in '/home/nelson/Workspace/outdoor_project_to_submit/region_detector/maskrcnn_benchmark/config/paths_catalog.py'
python3 ./tools/test_net.py --config-file '/home/nelson/Workspace/building_reconstruction/working_model/maskrcnn-boundary/configs/buildings_mask_rcnn_R_50_FPN_1x.yaml'

Ensembling primitives and relationships using IP

Set paths in run_ablation_experiments.py
Run python3 run_ablation_experiments.py

Citation

@inproceedings{nauata2020vectorizing,
  title={Vectorizing World Buildings: Planar Graph Reconstruction by Primitive Detection and Relationship Inference},
  author={Nauata, Nelson and Furukawa, Yasutaka},
  booktitle={European Conference on Computer Vision},
  pages={711--726},
  year={2020},
  organization={Springer}
}

yhexie/buildings2vec