Computer Vision Project

Project 2: Local Feature Matching

Example of a local features.

Overview

This project uses Matlab to programmatically match features via a modified SIFT pipeline. The project was fairly straightforward--the greatest challenge was working with Matlab as I still have very minimal experience with it.

Algorithm

The algorithm first detects interest points through a Harris corner detector. It then gathers features using SIFT by analyzing an image cell by cell. Finally, it matches features based on a ratio test. This has provided much better results than the original algorithm of simply cutting out basic image patches and comparing the two.

Harris Corner Detector

The Harris Corner Detector works in the following manner:

Blurring imag
First order derivative in x and y direction
Harris detector: (Ix^2 * Iy^2) - (Ixy^2) - alpha * (Ix^2+Iy^2)^2
Filter out values below threshold
Supress Edges
Find max values
Eliminate non-maxes

Local Feature Detection

The feature detection works in the following manner:

Calculate gradient magnitude and direction of Gaussian filtered image
Separate photo into grids
Calculate magnitude and direction of gradients within grid
Separate grids into cells
For each orientation, add the gradient magnitude sum to appropriate bin
Compile feature vector of results
Raise results to value less than one
Normalize results

Feature Matching

The feature matching works in the following manner:

Calculating the distance between the features
Sort by distance
Take ratio of the distances
Remove results greater than threshold
Return matching results

Results

For each image set, we show the visualization of the interest points dots, the visualization of potential matching features, and then the correctly matched features. Success rates in matches are listed. As discussed before, the success rate increased greatly when assessing for various orientaitons in the SIFT-like manner.