By Sizhe Hu, Wenming Wu, Ruolin Su, Wanni Hou, Liping Zheng and Benzhu Xu.
This repository is an official implementation of the paper Raster-to-Graph: Floorplan Recognition via Autoregressive Graph Prediction with an Attention Transformer. (EG 2024)
Raster-to-Graph is a novel automatic recognition framework, which achieves structural and semantic recognition of floorplans, addresses the problem of obtaining high-quality vectorized floorplans from rasterized images.
We represent vectorized floorplans as structural graphs embedded with floorplan semantics, thus transforming the floorplan recognition task into a structural graph prediction problem. We design an autoregressive prediction framework using the neural network architecture of the visual attention Transformer, iteratively predicting the wall junctions and wall segments of floorplans in the order of graph traversal. Additionally, we propose a large-scale floorplan dataset containing over 10,000 real-world residential floorplans. Extensive experiments demonstrate the effectiveness of our framework, showing significant improvements on all metrics. Qualitative and quantitative evaluations indicate that our framework outperforms existing state-of-the-art methods.
We contribute the following:
• An automatic recognition framework to obtain high-quality vectorized floorplans from rasterized images through one neural network.
• A novel autoregressive model that iteratively predicts structures and semantics of floorplans in the order of graph traversal.
• A large-scale floorplan dataset containing more than 10,000 realistic residential floorplans with dense annotations both on structures and semantics. To the best of our knowledge, this is currently the largest dataset available for floorplan recognition. The dataset has much potential to inspire more research.
To learn more, please refer to our paper.
The difference of about 0.1 in the Edge F-1 is because the 96.1 in the paper and the 96.2 here use different precision calculations (preserving 1 decimal place for Prec and Recall before calculation, and directly using the high-precision data inside the program)
Our repo was developed and tested with Python 3.7, cuda 11.1, Windows 10.
To install the necessary dependencies:
pip install -r requirements.txtor this:
conda create -n R2G python=3.7.13
pip install torch==1.9.1+cu111 torchvision==0.10.1+cu111 -f https://download.pytorch.org/whl/torch_stable.html
pip install cython==0.29.30 imageio==2.22.0 libpng==1.6.37 matplotlib==3.5.2 networkx==2.6.3 numpy==1.21.6 opencv-python==4.5.3.56 pillow==9.1.1 scikit-image==0.19.2 scipy==1.7.3 shapely==2.0.1 tqdm==4.64.0To install the necessary deformable attention module provided by Deformable-DETR:
cd models/ops/
sh make.sh
cd ../..(If you would like to know the version information of all packages, please refer to the requirements_full.txt.)
If you haven't received it by Monday morning (UTC+8), please check your spam folder. If you really haven't received it, please contact me: 2024010072@mail.hfut.edu.cn.
Our dataset includes:
(I) Centered 512*512 images obtained by processing "LIFULL HOME'S High Resolution Floor Plan Image Data" ("LIFULL HOME'S Data" for short).
We use centered 512*512 images (instead of "LIFULL HOME'S Data") in our model, because "LIFULL HOME'S Data" has considerable distinctions in sizes and margins.
(II) annotations.
Here is a detailed guide on how to get the data:
We do not have the permission to share the "LIFULL HOME'S Data". To access it, you need to apply through the following link: LIFULL HOME'S High Resolution Floor Plan Image Data. Upon approval, you will receive the images.
Their data contains from "photo-rent-madori-full-00" to "photo-rent-madori-full-0f", we only use "photo-rent-madori-full-00", so you do not need to download from "photo-rent-madori-full-01" to "photo-rent-madori-full-0f".
Now you should have a folder named "photo-rent-madori-full-00" that contains approximately 300000 images. You can place this folder at any path of your choice.
The annotations can be obtained from: Google Form Raster-to-Graph Dataset.
The download includes 3 folders:
annot_npy, annot_json: our annotations.
original_vector_boundary: boundary boxes of "LIFULL HOME'S Data". We use these data to process "LIFULL HOME'S Data" to centered 512*512 images. (Not used in our model.)
You simply need to place these 3 folders under the data folder.
Note: You need to use pillow=8.0.0 to run this script. Although requirements.txt specifies pillow=9.1.1, image_process.py was developed and tested with pillow=8.0.0. If you use pillow=9.1.1, the resulting processed images may have slight differences, but it shouldn't make a significant impact on the final testing results.
You should have gotten the "LIFULL HOME'S Data" in Step 1.
The 512*512 images are generated by running the image_process.py in data. Change the "original_images_path" variable to the path where you have placed the "photo-rent-madori-full-00" at, and just run the image_process.py.
After successfully running this script, the test, train, and val folders will contain 500, 9804, and 500 .jpg images, respectively, which will serve as the input for the model.
Your data directory should now contain image data (test, train, val each with 500, 9804, 500 images respectively) and annotation data (annot_npy with 10804 .npy files and annot_json with 3 .json files).
Here is a link to our trained model. To test on this model, please straightly run the following command:
python test.py
If you would like train your model, please adjust the arguments in args.py and then run:
python train.py
Note, at the 80th epoch, you should terminate the training process, modify:
(1) the lr=2e-4 lr_backbone=2e-5 lr_linear_proj=2e-5 to (2e-5 2e-6 2e-6) manually in args.py,
(2) the 'resume' in args.py to 'path/to/the/79th_checkpoint'
and rerun
To run the demo, you can prepare your own image folder (which can contain just 1 image), and then modify the demo.py file to change the input parameter of the MyDataset_demo. You can then run the following command:
python demo.py
The visualized output will be located in output/your_path_name. Note that the image data we used for training has been processed as described in the Data section earlier. If your image data is processed differently, you might need to retrain the model.
Please refer to guidance.docx.
If you use our code or dataset, please cite Raster-to-Graph:
Bibtex:
@article{https://doi.org/10.1111/cgf.15007,
author = {Hu, Sizhe and Wu, Wenming and Su, Ruolin and Hou, Wanni and Zheng, Liping and Xu, Benzhu},
year = {2024},
title = {Raster-to-Graph: Floorplan Recognition via Autoregressive Graph Prediction with an Attention Transformer},
journal = {Computer Graphics Forum},
volume = {43},
number = {2},
pages = {e15007},
keywords = {CCS Concepts, • Computing methodologies → Shape modeling, Computer vision},
doi = {https://doi.org/10.1111/cgf.15007},
url = {https://onlinelibrary.wiley.com/doi/abs/10.1111/cgf.15007},
eprint = {https://onlinelibrary.wiley.com/doi/pdf/10.1111/cgf.15007}
}
We thank Raster-to-Vector for their contribution to our data. According to Raster-to-Graph Terms of Use, if you use our dataset, we suggest that you also cite Raster-to-Vector:
Bibtex:
@INPROCEEDINGS{8237503,
author={Liu, Chen and Wu, Jiajun and Kohli, Pushmeet and Furukawa, Yasutaka},
booktitle={2017 IEEE International Conference on Computer Vision (ICCV)},
title={Raster-to-Vector: Revisiting Floorplan Transformation},
year={2017},
volume={},
number={},
pages={2214-2222},
keywords={Junctions;Semantics;Computational modeling;Solid modeling;Linear
programming;Three-dimensional displays;IP networks},
doi={10.1109/ICCV.2017.241}
}
If you have any questions, please contact me (Sizhe Hu) at 2024010072@mail.hfut.edu.cn.