This project implements a full image stitching pipeline using classical computer vision techniques:
- Harris Corner Detection
- SIFT Descriptor Extraction
- Descriptor Matching with Normalized Correlation & Euclidean Distance
- Match Filtering via Lowe’s Ratio Test
- Affine Transformation Estimation using RANSAC
- Warping and Overlay-based Stitching on a Shared Canvas
Developed in Python using OpenCV and NumPy.
Final evaluation includes qualitative results across diverse image categories and a parameter sensitivity analysis.
- 159 matches retained after ratio test @ 0.8
- All 159 confirmed as RANSAC inliers
- Avg. residual: 0.14 pixels
- Seamless visual alignment across cartoon-style architectural edges
🎨 Illustration used with credit to Maojin.Lee on ArtStation.
Displayed for non-commercial, educational purposes only.
- Retained matches: 398 after ratio test @ 0.8
- Avg. residual: 2.68 pixels
- Large terrain variation + weak keypoint correlation
- Failure due to parallax and repetitive textures
📷 Image source: Google Maps © imagery. Displayed for algorithm testing and educational use only.
- Trade-off identified around 0.7–0.8
- Lower ratio = cleaner matches, higher precision
- Higher ratio = more inliers, more residuals
| Image Pair | Matches | Inliers | Avg Residual | Outcome |
|---|---|---|---|---|
| Natural Scene | 248 | 247 | 1.01 px | ✅ Good Stitch |
| Sint-Pietersberg | 398 | — | 2.68 px | |
| Mount Fuji + Taranaki | 287 | — | 3.10 px | ❌ Poor Alignment |
| Van Gogh’s Starry Night | 500 | 500 | 0.00 px | ✅ Perfect |
| Google Maps | 124 | — | 2.69 px | ❌ Failed |
| Cartoon Buildings | 159 | 159 | 0.14 px | ✅ Excellent |
- Harris corner detector
- Threshold tuning @ 0.04–0.06
- Final threshold: 0.05
- Custom pipeline with SIFT descriptors
- Patch size: 5×5, 9×9, 15×15
- L2-normalization
- Cosine similarity & Euclidean distance
- Ratio test @ 0.6–0.9
- Final: 0.8
- Custom RANSAC (1000 iterations)
- Inliers: reprojection error < 3 px
- Residual: mean squared error across inliers
- Affine warp + canvas translation
- No blending (preserves edges)
- Matching fails under large color/texture variations
- SIFT not resilient to symbolic projections or cartoon styles
- Limited robustness to scale and orientation if too extreme
- Replace handcrafted features with learning-based descriptors (e.g. SuperPoint)
- Semantic-aware filtering of matches
- Homography or TPS warp instead of affine only
- Blending (feathering or Poisson) to improve visual continuity
image-stitching/
├── main.ipynb # End-to-end Harris + SIFT + RANSAC pipeline
├── results/ # Graphs and evaluation visualizations
├── figures/ # All stitched results and original input images
├── LICENSE
├── README.md # Project description and usage guide
└── requirements.txt # Python dependencies
Maintained by Hank Song
Originally developed as a Master's assignment at Maastricht University.