🧩 Panorama Stitching with OpenCV

A lightweight computer vision project that stitches multiple images into a single panorama. This implementation builds the full pipeline from scratch using feature matching, homography estimation, and blending techniques—without relying on high-level stitching APIs.

---

📌 Overview

This project constructs a multi-image stitching system that takes a sequence of overlapping images and produces a unified panoramic result.

Instead of using built-in stitching functions, the pipeline explicitly implements:

• Feature Detection (ORB)
• Feature Matching (KNN + Ratio Test)
• Robust Homography Estimation (RANSAC)
• sequential global alignment via accumulated pairwise homographies
• Distance-based Blending

---

🛠️ Key Features

1. Feature-based Alignment

• ORB is used for fast and efficient feature extraction
• KNN matching with Lowe’s Ratio Test helps improve matching reliability

2. Homography Estimation

• RANSAC removes outliers
• Inlier ratio is used to reject unreliable transformations (no fallback strategy)

3. Global Transformation

• A central reference frame is selected
• Homographies are accumulated left and right to align all images

4. Distance-based Blending

• Weighted blending using distance transform
• Reduces visible seams compared to naive averaging, but does not handle exposure or color differences

5. Dynamic Canvas Generation

• Canvas is estimated from warped boundaries with additional padding

---

📂 Project Structure

Image_stitching/
├── main.py                 # Main stitching pipeline
├── capture.py             # Image capture utility
├── images/                # Input images
└── panorama_result.jpg    # Output panorama

---

▶️ How to Run

python main.py

• Place input images in ./images/. or use capture.py to capture images from a webcam.

• Output:

  • panorama_result.jpg (full resolution)
  • Resized preview window

---

🖼️ Result

---

⚙️ Pipeline

Load Images
   ↓
Feature Extraction (ORB)
   ↓
Feature Matching (KNN + Ratio Test)
   ↓
Homography Estimation (RANSAC)
   ↓
Global Homography Accumulation
   ↓
Warping
   ↓
Distance-based Blending
   ↓
Final Panorama

---

🔍 Implementation Details

✔ Feature Matching

• Ratio test threshold: 0.7
• Top 100 matches used to reduce noise

✔ Homography Validation

inlier_ratio = np.sum(mask) / len(mask)
if inlier_ratio < 0.3:
    return None

✔ Blending

dist = cv.distanceTransform(mask, cv.DIST_L2, 5)
weight = dist / dist.max()

→ Pixels closer to the image center are weighted more heavily.

---

⚠️ Limitations & Analysis

Based on the implementation and observed results, several practical limitations were identified:

---

1. Parallax Effect (Major Issue)

If the camera undergoes translation instead of pure rotation:

• Near and far objects shift differently
• A single homography cannot model the scene

Result:

• Misalignment in overlapping regions
• Ghosting artifacts

---

2. Weak Features in Indoor Scenes

• Walls and ceilings lack texture
• Repetitive structures (e.g., bookshelves)

Result:

• Incorrect feature matches
• Unstable homography estimation

---

3. Blending Limitations

Current method:

• Distance-based feathering

Limitations:

• No exposure compensation
• Cannot fully eliminate seams with misalignment

---

4. Perspective Distortion

• Homography assumes planar scenes
• Real-world environments are 3D

Result:

• Warping distortion, especially at edges

---

5. Black Regions (Canvas Expansion)

• Warping introduces empty regions
• Cropping reduces final field of view

---

6. No Automatic Image Ordering

This implementation assumes that input images are already arranged in the correct sequential order.

• Images are stitched in a linear chain (i → i+1)
• No global matching or graph-based ordering is performed
• No similarity-based reordering is applied

Implications:

• Incorrect input order leads to:
  • Matching failure
  • Invalid homography estimation
  • Severe distortion or complete stitching breakdown

Requirement:

• Images must be captured and stored in order (e.g., left-to-right or right-to-left)
• File naming should reflect capture sequence

---

🚀 Possible Improvements

📷 Better Image Capture

• Keep camera position fixed
• Rotate around optical center
• Maintain 50–60% overlap

---

🧠 Algorithm Enhancements

✔ Multi-band Blending

• Separates low/high frequency components
• Produces seamless transitions

✔ Exposure Compensation

• Balances brightness across images

✔ Cylindrical Projection

• Reduces distortion for rotational panoramas

---

🔬 Advanced Techniques

• Bundle Adjustment (global optimization)
• Graph-based stitching
• Learning-based feature extraction (e.g., SuperPoint)

---

🧰 Tech Stack

• Python
• OpenCV
• NumPy