Floorplan Parser & Segmentation System

A deep learning-based system to parse architectural floorplans, extract structural layouts, and prepare them for simulation and analysis.

---

Features

• UNet-based floorplan segmentation
• Patch-based inference for large images
• Overlapping tile stitching (edge-aware)
• Accurate wall extraction from floorplans
• Handles large-scale layouts efficiently

---

Recent Improvements

• Adaptive stride tuning

  • Found optimal stride (~100) instead of fixed 128
  • Improves sampling diversity and reduces stitching artifacts

• Edge-aware padding (NEW)

  • Added ~5% reflective padding before inference
  • Eliminates boundary artifacts and improves corner predictions

• Improved stitching stability

  • Better overlap distribution across patches
  • Reduces repetition and seam artifacts

• Debug visualization

  • Raw probability maps (debug_raw_mask.png) added
  • Helps analyze model confidence before thresholding

Results

Original Floorplan

---

Raw Model Prediction (Before Fix)

---

Final Stitched Output (After Fix)

---

Debug of The Raw Masked Image

Problem Faced

When running segmentation on large floorplans:

• Model performed poorly on full-size images
• Edges were broken or missing
• Large layouts were not parsed correctly
• Context loss at boundaries

---

Solution: Patch-Based Segmentation + Smart Stitching

To fix this, we implemented a sliding window inference system with overlap and blending.

---

Step 1: Image Tiling

• Split large image into smaller patches (e.g., 256×256)
• Allows model to focus on local features

---

Step 2: Overlapping Inference

• Use stride < patch size

• Example: Patch Size = 256
  Stride = 128

• Each region is predicted multiple times

---

Step 3: Weighted Blending (Key Insight)

• Center of patch = high confidence
• Edges = low confidence

Weight map concept:

Center → Strong
Edges → Weak

This removes:

• seams
• edge artifacts
• broken walls

---

Step 4: Stitching

We combine all patch predictions using:

• weighted accumulation
• normalization

Final formula:

final_mask = sum(predictions * weights) / sum(weights)

---

Pipeline

        Input Image  
            V
   Add reflective padding (~5%)
            V
Split into overlapping patches  
            V  
      UNet Prediction  
            V  
     Weighted blending  
            V  
    Remove padding  
            V  
    Final stitched mask  

---

Project Structure

parser-model/
├── model.py
├── train.py
├── predict.py
├── predict_tiled.py
├── diagram_stitch.py
├── assets/
│ ├── original.png
│ ├── prediction.png
│ ├── stitched.png
│ ├── stitching_diagram.png
│ └── stitching.gif
├── README.md
├── LICENSE
└── .gitignore

---

python

python version used: 3.10.x

---

Dataset

This project was trained using the CubiCasa5K dataset.

CubiCasa5K is a large-scale annotated floorplan dataset containing detailed architectural layouts with semantic labels.

Due to size and licensing constraints, the dataset is not included in this repository.

You can access the dataset here:
https://github.com/CubiCasa/CubiCasa5k

If you use this dataset in your work, please consider citing the original authors.

---

Installation

pip install torch torchvision opencv-python numpy imageio

---

Requirements

 pip install -r requirements.txt

---

How to Run / Usage

python predict_tiled.py

---

Model Details

• Architecture: UNet
• Training data: ~1000 floorplan images
• Training epochs: 15
• Input resolution: 256 × 256

---

Limitations

• Model is trained on limited data and may not generalize to all architectural styles
• Fine details such as doors and furniture are not explicitly modeled
• Performance depends on preprocessing consistency

---

How Stitching Works

Large floorplans are processed using a sliding window approach with overlap.

Step-by-step:

1. The image is split into overlapping patches
2. Each patch is passed through the model
3. Predictions are combined using a weighted map
4. Overlapping regions are averaged to remove seams

---

Visualization

Overlapping patch concept for stitching

---

Stitching in Action

Sliding window stitching in action

The diagram explains how overlapping patches work, while the GIF shows the sliding window processing across the image in real time.

---

Why Overlap Matters

Without overlap:

• Broken edges
• Missing walls

With overlap:

• Smooth transitions
• Accurate structure

---

Important Notes

• Full-image inference is unreliable for large layouts
• Always use tiled inference for best results
• Preprocessing must match training pipeline
• Normalization is critical for correct predictions

---

Key Achievement

• Significantly improved large-scale floorplan segmentation
• Reduced edge artifacts using overlap + padding + blending
• Stabilized patch-based inference across different stride settings
• Achieved more consistent wall reconstruction on large layouts

---

Future Improvements

• Multi-scale inference
• Test-time augmentation
• Vectorization of wall structures
• Integration with evacuation simulation systems
• Detection of semantic elements (doors, exits)

---

License

This project is licensed under the MIT License.

---

Author

Sankhyapriyo Dey