pixel-stitch

From-scratch panoramic stitching - manual DLT, RANSAC, inverse warping & Laplacian pyramid blending. No cv2.Stitcher. Just math.

Panoramic Image Stitching — From Scratch

A complete panoramic image stitching pipeline built from scratch using only NumPy for computation and OpenCV for SIFT detection and image I/O. No high-level stitching APIs (e.g., cv2.Stitcher) are used.

Pipeline Overview

Input Images → SIFT Detection → Feature Matching → RANSAC Homography → Inverse Warping → Exposure Compensation → Multi-Band Blending → Crop → Panorama

Key Implementations (Manual)

Component	Method
Homography Estimation	Direct Linear Transform (DLT) with Hartley Normalization
Robust Estimation	RANSAC (10,000 iterations, 3.0 px threshold)
Image Warping	Inverse mapping with bilinear interpolation
Blending	Multi-band Laplacian pyramid (6 levels)
Exposure Compensation	Per-channel gain in overlap region
Cropping	Greedy shrink to largest clean rectangle

Permitted Library Usage

Library	Usage
OpenCV	Image I/O, SIFT detection, keypoint matching
NumPy	Matrix operations, SVD, linear algebra
Matplotlib	Visualization and figure export

---

File Structure

pixel-stitch/
├── Code/
│   ├── panorama_stitch.py              # Core stitching pipeline
│   └── panorama_stitch_memsafe.py      # Memory-hardened variant
├── Result/
│   ├── AppendixA/
│   │   ├── 01_input_images.png
│   │   ├── 02_sift_keypoints.png
│   │   ├── 03_matches_0_1.png
│   │   ├── 03_matches_2_1.png
|   |   └── output/
│   │       ├── 04_naive_stitch.jpg
│   │       ├── 04_naive_stitch_cropped.jpg
│   │       ├── 05_blended_panorama.jpg
│   │       └── 06_comparison.png
│   ├── AppendixB/
│   │   └── output/
│   │       └── ...
│   ├── AppendixC/                      # ← Required memsafe pipeline
│   │   └── output/
│   │       └── ...
│   ├── AppendixD/                      # ← Required memsafe pipeline
│   │   └── output/
│   │       └── ...
│   ├── AppendixE/                      # ← Required memsafe pipeline
│   │   └── output/
│   │       └── ...
│   ├── AppendixF/
│   │   └── output/
│   │       └── ...
|   ├── Compression.py
|   └── RESULTS.md
|
└── README.md

---

Usage

Core Pipeline

python panorama_stitch.py <left.jpg> <center.jpg> <right.jpg>

Supply at least 3 overlapping images ordered left to right. The center image is used as the reference frame (identity homography).

Memory-Safe Pipeline

python panorama_stitch_memsafe.py <left.jpg> <center.jpg> <right.jpg>

Use this variant when dealing with:

• High-resolution images (>4000 px per dimension)
• Scenes with repetitive structures (fences, blinds, tiled facades) that can cause degenerate homographies and massive canvas sizes
• Systems with limited RAM (<16 GB)

Differences Between the Two Scripts

Feature	panorama_stitch.py	panorama_stitch_memsafe.py
Core stitching logic	✓	✓
Multi-band Laplacian blending	✓	✓ (with memory guard)
float64 → uint8 compression	—	✓ (8× RAM savings)
Predictive OOM prevention	—	✓ (pre-calculates memory)
Disk-backed arrays (memmap)	—	✓ (for oversized canvases)
Linear blend fallback	—	✓ (when pyramids exceed RAM)

Requirements

Python >= 3.8
opencv-python >= 4.5
numpy >= 1.20
matplotlib >= 3.3

Install dependencies:

pip install opencv-python numpy matplotlib

Output Examples

The pipeline generates intermediate visualizations at each stage:

Output File	Description
01_input_images.png	Side-by-side view of all input images
02_sift_keypoints.png	SIFT keypoints overlaid on each image
03_matches_*.png	Feature correspondences between adjacent pairs
04_naive_stitch.png	Direct overlay with no blending (shows seams)
04_naive_stitch_cropped.png	Naive stitch after rectangular cropping
05_blended_panorama.png	Final panorama with multi-band blending
06_comparison.png	Naive vs. blended side-by-side comparison

Known Limitations

• Assumes pure rotational camera motion (no translational parallax). Imperfect rotation around the optical center will cause ghosting.
• Uses projective (planar) stitching, which introduces bow-tie distortion for wide-angle (>120°) panoramas. Cylindrical or spherical projection would be needed for ultra-wide mosaics.
• Exposure compensation uses a single per-channel gain — spatially varying illumination (vignetting) is not corrected.
• Scenes dominated by repetitive periodic structures can produce degenerate homographies; use panorama_stitch_memsafe.py for such cases.

License

This project was developed as a coursework assignment for Computer Vision.