Floor Segmentation using SVM

Project by: Adithya Kammati (230145)

---

Overview

This project implements a pixel-wise floor segmentation model using Support Vector Machine (SVM) with rich multi-scale features. The model classifies each pixel in an image as either "floor" or "non-floor".

Final Results

• Accuracy: 96.75%
• Precision (Floor): 96%
• Recall (Floor): 98%
• F1-Score: 97%

---

Dataset

• Location: adithya_230145/ folder in Google Drive
• Images: 10 floor images in various formats (jpg, jpeg, webp)
• Annotations: COCO JSON format polygon masks (coco_annotations.json)
• Total Samples: 100,000 pixels (50K floor + 50K non-floor)

---

Features Extracted (20 per pixel)

Feature	Description
RGB (3)	Red, Green, Blue color values
HSV (3)	Hue, Saturation, Value color space
Grayscale (1)	Intensity value
Multi-scale LBP (3)	Local Binary Pattern at R=1,2,3
Gaussian Blur (2)	Smoothed versions at sigma=2,5
Local Stats (2)	Patch mean and standard deviation
Edge Magnitude (1)	Sobel edge strength
Spatial Position (2)	Normalized x,y coordinates
DoG-like (1)	Difference of Gaussians
Color Ratios (2)	R-B difference, G/R ratio

---

Model Architecture

• Algorithm: Support Vector Machine (SVM)
• Kernel: RBF (Radial Basis Function)
• Parameters: C=50, gamma=0.1
• Scaler: StandardScaler (z-score normalization)

---

How to Use the Trained Model

import joblib
import cv2
import numpy as np
from skimage.feature import local_binary_pattern
from scipy.ndimage import uniform_filter, gaussian_filter

# Load model and scaler
clf = joblib.load("svm_floor_model_v2.pkl")
scaler = joblib.load("svm_scaler_v2.pkl")

# Extract features from a new image
def extract_features_for_prediction(image):
    # (Same feature extraction as training)
    # Returns feature array for all pixels
    pass

# Predict
features = extract_features_for_prediction(new_image)
features_scaled = scaler.transform(features)
prediction = clf.predict(features_scaled)
mask = prediction.reshape(image_height, image_width)

---

Files Saved

File	Description
svm_floor_model_v2.pkl	Trained SVM model
svm_scaler_v2.pkl	Feature scaler

---

Improvement from Baseline

Version	Features	Accuracy
v1 (Basic)	4 features (LBP + RGB)	73.47%
v2 (Improved)	20 features (multi-scale)	96.75%

Improvement: +23.28%

---

Future Improvements

1. Use deep learning (SegFormer, DeepLabV3+) for 98-99% IoU
2. Add more training images (500+)
3. Use SAM for auto-labeling
4. Implement CRF post-processing

---

References

• scikit-learn SVM documentation
• Local Binary Patterns (LBP) for texture
• COCO dataset format