| id |
ecaade2024_262 |
| authors |
Guo, Yu; Cui, Zhe |
| year |
2024 |
| title |
Impact of the Training Set Consistency on Architectural Plan Generation Effect Based on Pix2pixHD Algorithm |
| doi |
https://doi.org/10.52842/conf.ecaade.2024.1.459
|
| source |
Kontovourkis, O, Phocas, MC and Wurzer, G (eds.), Data-Driven Intelligence - Proceedings of the 42nd Conference on Education and Research in Computer Aided Architectural Design in Europe (eCAADe 2024), Nicosia, 11-13 September 2024, Volume 1, pp. 459–468 |
| summary |
In research on machine learning-assisted architectural plan generation, the sample composition of the training set is one of the most important factors influencing the model's performance and outcomes. In most previous research, architectural plans in the training set exhibited a lack of consistency in design principles, which hindered generation effectiveness. In this study, we have developed an 'architecture-like plan' dataset that adheres to a set of unified principles. We carried out a training-testing experiment based on the pix2pixHD algorithm with architecture-like plans and quantitatively evaluated the similarity between the predictions and ground truths by the pixelmatch algorithm. The similarity was high, up to 92.26%, and the predictions were high quality, suggesting that the algorithm has learned the design principle. The result significantly outperforms similar studies, suggesting that the training set consistency positively affects the generation effect. Next, we validate this on the open-source residential plan dataset (RPLAN) through another training-testing experiment. We filtered a subset with uniform criteria, containing design principles and label accuracy, as the experimental group, and the original unfiltered dataset as the control group. The results showed that the similarity of the experimental group achieved 72.03%, compared to the control group's (55.13%), and the experimental group's predictions were significantly superior to those of the control group. Both experiments show that the higher the training set consistency, the more likely it is to obtain a generative model with excellent results, and that the training set consistency significantly affects the generation of architectural plans. |
| keywords |
training set consistency, architectural plan generation, machine learning, pix2pixHD |
| series |
eCAADe |
| email |
|
| full text |
 file.pdf (1,165,750 bytes) |
| references |
Content-type: text/plain
|
| last changed |
2024/11/17 22:05 |
|
