---

FloorPlan2Design

Overview

Aigc project during my sophomore year internship at the Institute of Automation, Chinese Academy of Sciences

FloorPlan2Design is a project I completed during my internship, focusing on automatically converting building floor plans into detailed design renderings. After reading the ArchiGAN paper, I used the pix2pix conditional GAN ​​(CGAN) framework, which, while not state-of-the-art, is a solid foundation for this image-to-image translation task.

Main Features

• Automatic Translation: Convert building floor plans into design renderings with minimal human intervention.

• Proven Framework: Uses pix2pix CGAN as the image-to-image translation model.

• Customizable Data Pipeline: Includes extensive preprocessing and data management tools to enhance model performance.

Dataset

In order to avoid too much redundancy in downloading data, I only left 4 to 6 images of all the data for demonstration and file location prompts.

I used the CVC-FP dataset for training. This dataset is of very poor quality, but there are no good dataset resources, so I can only do data cleaning and data processing on this dataset.

After downloading the dataset, the file should be placed in the ImagesGT folder. During the preprocessing and training phase, several problems were found and solved:

• [x] SVG rendering problem: Due to code corruption, most SVG annotations cannot be rendered in browsers or other viewing tools.

• [x] Room classification: The original SVG file does not clearly classify room types, but uses a general "room" label, which increases the difficulty of distinguishing room types.

• [x] File naming mismatch: Some images and their corresponding SVG files do not match in name or extension.

• [x] Unmatched SVGs: There are unrelated images without corresponding SVG files.

• [x] Duplicate images: Some images are duplicated and unusable.

• [x] Inconsistent color codes: The color codes in the SVG annotations vary from file to file.

The raw dataset posed significant challenges and required extensive preprocessing. I developed a batch processing script to correct and standardize all SVG annotations. The Python script is thoroughly documented to help understand the functionality of each component.

Run the project

Just download the project and the datasets mentioned above, find the python files in the utils repository that process the datasets, because each file has very clear comments, and use ipynb to run the project after processing all the dirty data. I wrote 3 models for this project to complete the task, ensuring that the internship is successfully completed and each project can run the results.

Visual results

Here are some examples of visual output generated by the model:

Folder structure

Here is an overview of the project structure:

  

═── Dataset # Dataset used in the project

│ ═── A # Black and white outline image

│ ═── B # Image with color annotations

│ ═── ImagesGT # Original downloaded image

│ ═── Initial_Data

│ │ ═── colour # Image with updated color scheme

│ │ │ ═── 1_windows

│ │ │ ═── 2_all_room

│ │ │ ═── 3_all_wall

│ │ │ ═── 4_kitchen_room

│ │ │ ═── 5_livingroom_room

│ │ │ ═── 6_bathroom_room

│ │ │ ═── 7_bedroom_room

│ │ │ └── final

│ │ ═── final_svg # Final, corrected SVG file

│ │ ═── fix_svg # Temporary SVG correction

│ │ ═── PNG # Converted PNG file

│ │ └── SVG # Original SVG file

│ └── svg_a_b # Final usable SVG data

│ ═── a

│ └── b

═── Training

│ ═── checkpoints # Model checkpoints during training

│ ═── generated_images # Images generated during training

│ │ ═── 1

│ │ ═── 2

│ │ └── 3

│ └── logs # Training logs

└── utils # Utility scripts for SVG modification

└── colour_change # Tool for changing SVG color scheme

Getting Started

1. Dataset Preparation:

• Download the CVC-FP dataset and place the images in the ImagesGT directory.

• Run the preprocessing scripts to correct and normalize the dataset.

2. Model Training:

• Follow the instructions in the Training directory to start training the model.

• Checkpoints and logs will be saved automatically.

Acknowledgements

Special thanks to the Institute of Automation, Chinese Academy of Sciences for providing the project internship opportunity.

---