This is an implementation of SIFT (David G. Lowe's scale-invariant feature transform) done entirely in Python with the help of NumPy. This implementation is based on OpenCV's implementation and returns OpenCV
KeyPoint objects and descriptors, and so can be used as a drop-in replacement for OpenCV SIFT. This repository is intended to help computer vision enthusiasts learn about the details behind SIFT.
PythonSIFT has been reimplemented (and greatly improved!) in Python 3. You can find the original Python 2 version in the
legacy branch. However, I strongly recommend you use
master (the new Python 3 implementation). It's much better.
Last tested successfully using
Numpy 1.19.4 and
import cv2 import pysift image = cv2.imread('your_image.png', 0) keypoints, descriptors = pysift.computeKeypointsAndDescriptors(image)
It's as simple as that. Just like OpenCV.
keypoints are a list of OpenCV
KeyPoint objects, and the corresponding
descriptors are a list of
128 element NumPy vectors. They can be used just like the objects returned by OpenCV-Python's SIFT
detectAndCompute member function. Note that this code is not optimized for speed, but rather designed for clarity and ease of understanding, so it will take a few minutes to run on most images.
You can find a step-by-step, detailed explanation of the code in this repo in my two-part tutorial:
I'll walk you through each function, printing and plotting things along the way to develop a solid understanding of SIFT and its implementation details.
I've adapted OpenCV's SIFT template matching demo to use PythonSIFT instead. The OpenCV images used in the demo are included in this repo for your convenience.
Anyone is welcome to report and/or fix any bugs. I will resolve any opened issues as soon as possible.
Any questions about the implementation, no matter how simple, are welcome. I will patiently explain my code to you.
Definitely worth a read!
SIFT was patented, but it has expired. This repo is primarily meant for educational purposes, but feel free to use my code any way you want, commercial or otherwise. All I ask is that you cite or share this repo.
You can find the original (now expired) patent here (Inventor: David G. Lowe. Assignee: University of British Columbia.).