PythonSIFT

PythonSIFT is an educational, pure-Python implementation of David G. Lowe's seminal Scale-Invariant Feature Transform (SIFT) algorithm. Built entirely on top of NumPy, it is explicitly designed to mirror the behavior and data structures of OpenCV's C++ SIFT implementation while remaining accessible, readable, and highly customizable.

The Philosophy Behind PythonSIFT

While highly optimized, production-grade C++ implementations of SIFT exist (most notably within the OpenCV library), their codebases are often heavily abstracted, heavily macro-driven, and extremely difficult for computer vision students or researchers to parse.

PythonSIFT prioritizes algorithmic clarity, mathematical transparency, and ease of understanding over execution speed.

By leveraging this repository, developers and computer vision enthusiasts can:

  • Step through the algorithm line-by-line in a Python debugger.
  • Inspect intermediate matrix states, such as the intricate Difference-of-Gaussians (DoG) image pyramid.
  • Experiment with core logic, such as modifying the quadratic fit used for sub-pixel keypoint localization or altering the orientation histogram binning.

Core Capabilities and Features

  • Pure Python & NumPy Architecture: No compiled C/C++ extensions, Cython, or complex build tools are required. The entire algorithmic pipeline relies purely on standard scientific Python operations.
  • 1:1 OpenCV Compatibility: PythonSIFT acts as a drop-in replacement for cv2.SIFT_create().detectAndCompute(). It returns standard OpenCV KeyPoint objects and float32 NumPy array descriptors. This allows seamless integration into existing OpenCV matching pipelines (e.g., using cv2.FlannBasedMatcher).
  • Educational Modularity: The code is meticulously organized into the distinct stages described in Lowe's original paper: Scale-space extrema detection, Keypoint localization, Orientation assignment, and Descriptor generation.

When to Use PythonSIFT (and When Not To)

Use PythonSIFT when:

  • You are studying computer vision and want to understand the intricate mathematics behind feature extraction.
  • You are prototyping a custom variant of SIFT (e.g., changing the descriptor size, testing new edge suppression techniques).
  • You need to extract features in an environment where compiling OpenCV from source with non-free modules (historically an issue with SIFT) is restrictive, and performance is secondary.

Do NOT use PythonSIFT when:

  • You are building a production application requiring real-time performance (e.g., live video tracking, SLAM). PythonSIFT is not optimized for speed and will take significantly longer to process an image than OpenCV's C++ implementation. For production, always use cv2.SIFT_create().

Important Notice: Naming Confusion

Critical Warning: Please note that this project is called PythonSIFT, not PySift.

There is an unrelated, broken package on PyPI registered under the name pysift that frequently crashes with urlparse errors. Do not run pip install pysift expecting to get this repository. Before opening an issue, ensure you are actually using the source code cloned directly from this GitHub repository. See the Installation Guide for proper setup instructions.

SIFT was previously protected by a strict software patent (US6711293B1, Inventor: David G. Lowe, Assignee: University of British Columbia). Because of this, it was excluded from standard OpenCV distributions for years.

However, the SIFT patent has officially expired. This repository is intended for educational purposes, but the code may be used freely (commercial or otherwise) under the terms of the MIT License. If you find this repository helpful in your academic or professional work, citing or sharing the repository is highly appreciated.