Image Stitching with OpenCV

A clean, notebook-based computer vision project that stitches two overlapping images into a single panorama using classical feature matching and geometric transformation techniques.

Overview

This project demonstrates a full panorama stitching workflow implemented in OpenCV inside a Jupyter notebook. The pipeline loads a left and right image, detects local features with SIFT, matches descriptors with BFMatcher, estimates a robust projective transform using RANSAC, warps one image into the other image's coordinate space, and trims the final panorama to remove empty borders.

The repository is a good fit for:

• computer vision coursework,
• OpenCV practice,
• homography and panorama demos,
• portfolio projects that show end-to-end image alignment.

Pipeline

The notebook follows this sequence:

1. Load input images
   Reads left.jpg and right.jpg, then converts them to grayscale for feature extraction.

2. Detect keypoints and descriptors
   Uses cv2.xfeatures2d.SIFT_create() to extract scale-invariant keypoints and descriptors.

3. Match descriptors
   Uses cv2.BFMatcher() with knnMatch(..., k=2) to retrieve the two nearest matches per descriptor.

4. Filter matches
   Applies Lowe-style ratio filtering to keep stronger correspondences.

5. Estimate homography
   Uses cv2.findHomography(..., cv2.RANSAC, 5.0) to compute the perspective transformation between the two views.

6. Warp and stitch
   Uses cv2.warpPerspective(...) to project one image into panorama space, then copies the second image into the canvas.

7. Trim borders
   Removes black borders from the stitched output with a recursive trimming function.

Repository Contents

.
├── Image-stiching.ipynb   # Main notebook (original filename kept as-is)
├── left.jpg               # Input image A (expected)
├── right.jpg              # Input image B (expected)
└── README.md              # Project documentation

Requirements

• Python 3.7+
• Jupyter Notebook or JupyterLab
• OpenCV with contrib modules
• NumPy
• Matplotlib

Install dependencies with:

pip install opencv-contrib-python numpy matplotlib jupyter

How to Run

1. Clone the repository.
2. Place two overlapping images in the project folder named:
   • left.jpg
   • right.jpg
3. Start Jupyter:

jupyter notebook

4. Open Image-stiching.ipynb.
5. Run the notebook cells from top to bottom.

Generated Outputs

The notebook creates these files during execution:

• keypoints.jpg — detected keypoints visualization.
• warped.jpg — perspective-warped image.
• output.jpg — initial stitched panorama before trimming.
• finaloutput.jpg — trimmed final panorama.

Example Code Sections

SIFT feature extraction

sift = cv2.xfeatures2d.SIFT_create()
kp1, des1 = sift.detectAndCompute(img1, None)
kp2, des2 = sift.detectAndCompute(img2, None)

Descriptor matching

match = cv2.BFMatcher()
matches = match.knnMatch(des1, des2, k=2)

Homography estimation

H, masked = cv2.findHomography(src, dst, cv2.RANSAC, 5.0)

Perspective warping

dst = cv2.warpPerspective(img_, H, (img.shape[1] + img_.shape[1], img.shape[0]))

Notes and Limitations

• The notebook depends on SIFT, which requires the OpenCV contrib package.
• Stitching quality depends heavily on overlap, texture, and lighting consistency between the two images.
• The current implementation is designed for two images only.
• Blending is minimal; visible seams can still appear in difficult cases.

Improvement Ideas

Possible upgrades for a stronger production version:

• Add automatic canvas sizing and smarter cropping.
• Support more than two images.

Poor stitching result

Check that:

• the images have enough overlap,
• the scene is mostly static,
• the images were taken from roughly the same optical center,
• lighting changes are not too strong.