PanoStitch

A modular panoramic image stitching pipeline featuring both classic algorithms and modern Deep Learning integrations.

Overview

PanoStitch is an intelligent panoramic image stitching system that seamlessly combines multiple overlapping images into a single high-resolution panorama. Our modular architecture supports both traditional computer vision techniques and cutting-edge deep learning approaches, offering flexibility and robustness for various stitching scenarios.

Input Images

| Image 1 | Image 2 | | -------------------------------------- | --------------------------------------------- | | | |

Result Panorama

Pipeline Architecture

flowchart TD
    A[Input Images] --> B[Feature Detection]
    B --> C[Feature Description]
    C --> D[Feature Matching]
    D --> E[Homography Estimation]
    E --> F[Gain Compensation]
    F --> G[Image Blending]
    G --> H[Output Panorama]
    
    style A fill:#e3f2fd,stroke:#1976d2
    style B fill:#f3e5f5,stroke:#7b1fa2
    style C fill:#fce4ec,stroke:#c2185b
    style D fill:#fff3e0,stroke:#f57c00
    style E fill:#e8f5e9,stroke:#388e3c
    style F fill:#e0f7fa,stroke:#00838f
    style G fill:#fff8e1,stroke:#ffa000
    style H fill:#c8e6c9,stroke:#2e7d32

Test Cases: You can find 8 comprehensive test cases with their stitched panorama results in the imgs/ folder (bicycle, boat, clock, dam, flower, mountain, river, tree).

Features

• Feature Detection: Custom Harris Corner Detector or SIFT.
• Feature Description: Custom 128D HOG-based descriptors.
• Matching: Vectorized Brute-force matcher with Lowe's ratio test.
• Deep Learning: Integration with DISK + LightGlue for robust matching.
• Homography: RANSAC with Direct Linear Transform (DLT).
• Processing: Gain compensation for exposure correction and weighted blending.

Installation

pip install -r requirements.txt

Usage

Basic Usage (Recommended)

Stitch all images in a directory using the standard SIFT pipeline:

python panostitch.py imgs/boat/

Comparing Methods

1. Custom "From Scratch" Implementation (Harris + HOG Descriptors)

python panostitch.py imgs/boat/ --harris

2. Deep Learning Pipeline (DISK + LightGlue)

python panostitch.py imgs/boat/ --dnn

Benchmark Comparison

We benchmarked three matching methods across 8 scenes: traditional SIFT, and deep learning approaches LoFTR and DISK+LightGlue.

Feature Matching Visualization

| SIFT (Traditional) | DISK+LightGlue (Deep Learning) | LoFTR (Deep Learning) | |--------------------|-------------------------------|-----------------------| | | | |

Speed Comparison

| Method | Avg Time | Min | Max | |--------|----------|-----|-----| | SIFT | 2.0s | 0.46s | 7.92s | | DISK+LightGlue | 14.8s | 12.9s | 16.3s | | LoFTR | 30.1s | 25.2s | 32.5s |

Match Quality by Scene

| Scene | Method | Matches | Inliers | Inlier Ratio | |-------|--------|---------|---------|--------------| | boat | SIFT | 3836 | 3227 | 84.1% | | | DISK+LightGlue | 1238 | 1209 | 97.7% | | | LoFTR | 4503 | 1963 | 43.6% | | dam | SIFT | 3883 | 3856 | 99.3% | | | DISK+LightGlue | 1319 | 1272 | 96.4% | | | LoFTR | 4218 | 4213 | 99.9% | | mountain | SIFT | 1277 | 1267 | 99.2% | | | DISK+LightGlue | 1190 | 1174 | 98.7% | | | LoFTR | 4994 | 4495 | 90.0% | | clock | SIFT | 1626 | 1389 | 85.4% | | | DISK+LightGlue | 592 | 517 | 87.3% | | | LoFTR | 4214 | 1541 | 36.6% | | bicycle | SIFT | 774 | 558 | 72.1% | | | DISK+LightGlue | 824 | 707 | 85.8% | | | LoFTR | 4606 | 1826 | 39.6% | | flower | SIFT | 591 | 433 | 73.3% | | | DISK+LightGlue | 1097 | 893 | 81.4% | | | LoFTR | 4060 | 1013 | 25.0% | | tree | SIFT | 570 | 301 | 52.8% | | | DISK+LightGlue | 214 | 117 | 54.7% | | | LoFTR | 3995 | 664 | 16.6% | | river | SIFT | 2178 | 2139 | 98.2% | | | DISK+LightGlue | 898 | 893 | 99.4% | | | LoFTR | 4675 | 4657 | 99.6% |

Key Findings

graph TB
    subgraph "Speed"
        A["SIFT: 2.0s avg<br/>7-15x Faster"]
    end
    subgraph "Quality"
        B["DISK+LightGlue<br/>85-99% Inlier Ratio"]
    end
    subgraph "Coverage"
        C["LoFTR<br/>4000-5000 matches"]
    end
    
    style A fill:#e8f5e9
    style B fill:#fff3e0
    style C fill:#e3f2fd

Summary:

• SIFT is 7-15x faster than deep learning methods, ideal for real-time applications
• DISK+LightGlue achieves highest inlier ratios (85-99%) with moderate match counts
• LoFTR produces the most matches but lower inlier ratios in challenging scenes
• On well-textured scenes (dam, river), all methods achieve >96% inlier ratio

Contributors

| <a href="https://avatars.githubusercontent.com/AmiraKhalid04?v=4"><img src="https://avatars.githubusercontent.com/AmiraKhalid04?v=4" alt="Amira Khalid" width="150"></a> | <a href="https://avatars.githubusercontent.com/u/121557650?v=4"><img src="https://avatars.githubusercontent.com/u/121557650?v=4" alt="Youssef Noser" width="150"></a> | <a href="https://avatars.githubusercontent.com/u/149144312?v=4"><img src="https://avatars.githubusercontent.com/u/149144312?v=4" alt="Hussein" width="150"></a> | <a href="https://avatars.githubusercontent.com/u/136710727?v=4"><img src="https://avatars.githubusercontent.com/u/136710727?v=4" alt="Abdulrahman Medhat" width="150"></a> | | :----------------------------------------------------------------------------------------------------------------------------------------------------------------------: | :-------------------------------------------------------------------------------------------------------------------------------------------------------------------: | :-------------------------------------------------------------------------------------------------------------------------------------------------------------: | :------------------------------------------------------------------------------------------------------------------------------------------------------------------------: | | Amira Khalid | Youssef Noser | Hussein Mohammed | Abdulrahman Medhat |