Foreground Segmentation and Poisson Blending

This project implements the GrabCut algorithm for foreground segmentation and utilizes Poisson Blending for background replacement. The goal is to accurately segment a foreground object from an image and seamlessly blend it with a new background, preserving texture and color.

Introduction

This repository contains the implementation of both GrabCut and Poisson Blending algorithms, including code, example images, and a detailed report on the methodology and results.

Methodology

GrabCut Algorithm

1. Initialization of GMMs:

   • Initialize Gaussian Mixture Models (GMMs) for foreground and background using K-means clustering.

2. Update GMMs:

   • Calculate and update mean, covariance, and weights of GMM components based on the current mask.

3. Min-Cut Calculation:

   • Build a graph using the mask and energy terms, then compute the min-cut to segment the image.

4. Mask Update:

   • Update the mask based on the min-cut results.

5. Convergence Check:

   • Determine if the energy value has stabilized to a minimum.

6. Evaluation Metric:

   • Evaluate segmentation accuracy and Jaccard similarity against ground truth.

Example Images

Poisson Blending

1. Laplacian Operator Calculation:

   • Compute the Laplacian of the source image.

2. Solve Poisson Equation:

   • Solve the Poisson equation to blend the source and target images.

3. Blend and Save Image:

   • Blend the foreground object with the new background and save the result.

Example Images

Detailed Report

A comprehensive report on the methodology and results can be found here.

Repository Structure

data/imgs/
  ├── ... (input images)
data/seg_GT/
  ├── ... (ground truth binary images)
data/bg/
  ├── ... (background images)
data/bboxes/
  ├── ... (bounding boxes as txt files)
report_examples/
 ├── ... (example images and a detailed report)
grabcut.py
poisson_blend.py
README.md