CS 4476/6476 Project 2: SIFT Local Feature Matching

🧠 SIFT-Based Local Feature Matching & Panorama Stitching

📌 Project Overview

This project implements a SIFT-inspired feature matching pipeline for matching images taken from multiple viewpoints of the same scene. The pipeline includes:

• Harris corner detection
• Local patch and SIFT-like descriptors
• Ratio-test-based feature matching
• Panorama stitching using estimated homographies

🚀 Features Implemented

🧭 1. Harris Corner Detector

• Uses image gradients (Sobel filter) and second-moment matrix
• Computes Harris response score:
  $$ R = \text{det}(A) - \alpha \cdot (\text{trace}(A))^2 $$
• Applies max-pooling for non-maximum suppression (NMS)

🔲 2. Normalized Patch Descriptor

• Extracts fixed-size grayscale patches centered at interest points
• Normalizes each patch to unit norm for simple but effective matching
• Expected matching accuracy on Notre Dame: ~40–50%

📌 3. Ratio Test Feature Matching

• Implements Lowe's Nearest Neighbor Distance Ratio (NNDR) test: $$ \text{Ratio} = \frac{||f_1 - f_{match1}||}{||f_1 - f_{match2}||} $$
• Filters robust matches by thresholding this ratio
• Returns top confident matches and visualizes them

🌀 4. SIFT-Like Descriptor

• Computes 4×4 grid of 8-bin gradient histograms over 16×16 patch
• Applies Gaussian weighting and orientation binning
• Performs Square Root SIFT normalization for robustness
• Achieves >80% accuracy on Notre Dame image pair

🌄 5. Panorama Stitching (Hand-Graded)

• Detects features and estimates homography using cv2.findHomography()
• Warps and blends images using cv2.warpPerspective()
• Stitching tested on 3 given panoramas + 1 custom pair

Getting started

• See Project 0 for detailed environment setup.
• Ensure that you are using the environment cv_proj2, which you can install using the install script conda/install.sh.