GSDiff

Official implementation of the AAAI 2025 paper: "GSDiff: Synthesizing Vector Floorplans via Geometry-enhanced Structural Graph Generation"

Data

1. Create folder datasets/rplandata/Data.

2. Download the 80,788 RPLAN dataset (http://staff.ustc.edu.cn/~fuxm/projects/DeepLayout/index.html). It contains a floorplan_dataset folder. Place this floorplan_dataset folder under datasets/rplandata/Data.

3. Run the following scripts to obtain structural graph data:

   python rplan-extract.py
   python rplan-process1.py
   python rplan-process2.py
   python rplan-process3.py
   python rplan-process4.py

   • After completion, a directory rplang-v3-withsemantics (65,763 train + 3,000 val + 3,000 test = 71,763 .npy files) will be created under datasets/rplandata/Data.
   • Running rplan-extract.py also generates folders (1-channel-semantics-256, 3-channel-semantics-256, bin_imgs, e_imgs, etc.). You can remove them.

4. Run the following scripts to obtain structural graph data with boundaries:

   python rplan-process5.py
   python rplan-process6.py
   python rplan-process7.py

   • After completion, 2 directories rplang-v3-withsemantics-withboundary and rplang-v3-withsemantics-withboundary-v2, will appear under datasets/rplandata/Data, each containing 71,763 files.

5. Run the following scripts to obtain topology graphs:

   python rplan-process8.py
   python rplan-process9.py
   python rplan-process10.py

   • After completion, a rplang-v3-bubble-diagram folder will be created under datasets/rplandata/Data, containing the same number of files.

6. Move data from datasets/rplandata/Data to datasets for train/val/testing:

   python move.py

Note, when we conducted our experiments, the semantics of bubble diagram GT involved randomness. Due to the terms of the RPLAN dataset, we are not permitted to release any part of it. Therefore, the bubble diagram GT semantics extracted using the provided scripts may differ slightly from our experimental results. However, given the large data scale, the bias should be minor (for rooms with ambiguous categories, both our GT and the GT you extract randomly select one category, which is no statistically significant difference). Alternatively, you can use get_cycle_basis_and_semantic_3_semansimplified instead of get_cycle_basis_and_semantic_2_semansimplified in rplan-process8/9/10.py to extract room semantics and train your own topology models. This method is not random, may yield improvements over the metrics reported in the paper.

--------------------------LIFULL---------------------------------------

If you want to try training/generating on the LIFULL dataset, please create path datasets/lifulldata and follow the data request process of Raster-to-Graph (https://github.com/SizheHu/Raster-to-Graph) to place the data under this path datasets/lifulldata.

The data contains 10,804 images (Step 1: Access the "LIFULL HOME'S Data") and corresponding annotations (Step 2: Access the Annotations).

Usage

The test scripts for no constraints, topology constraints, and boundary constraints are all placed under scripts (test_xxx.py). Download the corresponding weights and run them via:

python test_xxx.py

No constraints: We use the original 3000 results and run them 5 times to get the average.

Topology constraints: We took the intersection of the original 3000 results with the test set numbers of HouseDiffusion and House-GAN++, and got 757. We ran them 5 times and averaged them to get the FID, KID, GED, and statistical analysis of each room type. The sample numbers of 757 are in line 183 of evalmetric-topoconstrain-ged-roomnumber.py.

Boundary constraints: We took the intersection of the original 3000 results with the test set numbers of HouseDiffusion and House-GAN++, and got 378. We ran them 5 times and averaged them to get the FID, KID, GED, and statistical analysis of each room type. The sample number of 378 is on line 9 of evalmetric-boun-constrain-fid-kid.py.

--------------------------LIFULL---------------------------------------

All training and testing scripts on LIFULL dataset have 'lifull' in the file names.

Like RPLAN dataset, the purpose of each script is stated at the top of the script.

params (place in the 'outputs' folder)

unconstrained params: https://drive.google.com/file/d/15gM0GtW2GwHmlpz0r-rpvo-k-BlNy_gu/view?usp=sharing

topology-constrained params: https://drive.google.com/file/d/1pk7SmvLZ8ON3OUL3SNxPRu73ndVKru0z/view?usp=sharing

boundary-constrained params: https://drive.google.com/file/d/1puqxXIW4Y7AeQHFuC76PlYpWQm6MD8PS/view?usp=sharing

boundary-autoencoder CNN params: https://drive.google.com/file/d/1l6QRpfX5Jtucg3R995HajlwRG8SewUJW/view?usp=sharing

topology-autoencoder Transformer params: https://drive.google.com/file/d/1tExX8LdrFpJfBQH5y2emC6BltBwf9tHx/view?usp=sharing

unconstrained sloping walls params: https://drive.google.com/file/d/1aQNaQwHPkdlNyMQoazZDye7_O8Qbqvzq/view?usp=drive_link

--------------------------LIFULL---------------------------------------

Training parameters on the LIFULL dataset:

Node: https://drive.google.com/file/d/1k_q9-vQXbs3PDzLxvz-tQRvO3j0DzlPN/view?usp=sharing

Edge: https://drive.google.com/file/d/1XkoMZAMOeBPTteUTVDukgc4BNoEEJSXS/view?usp=sharing