PyVision Computer Vision Toolkit

by David Bolme

As of August 2014, PyVision is now hosted at github: http://github.com/bolme/pyvision

For additional documentation please see the github wiki: http://github.com/bolme/pyvision/wiki

PyVision is a object-oriented Computer Vision Toolkit (BSD Licenses) that is designed to allow the rapid prototype and analysis of computer vision algorithms. Using python as a foundation this package provides a simple framework that unifies the Matlab like functionality of scipy/numpy, Open Computer Vision Library (opencv), and other vision and machine learning software packages. In addition PyVision provides a set of analysis tools that allows the researcher to evaluate there algorithms in python or to export those results in csv format for analysis in Excel, R, or SAS.

PyVision is a collection of computer vision algorithms that I have implemented as part of my Ph.D. face recognition research. The purpose of the library is to provide a set of utilities that interface with common computer vision and machine learning libraries that will allow rapid vision algorithm prototyping. For example the current implementation has an image class which will easily translate image data between PIL, numpy array, and opencv formats.

Overview

Currently, PyVision contains a set of popular computer vision algorithms including:

• canny edge detection
• harris corner detection
• viola and jones face detection
• support vector machines and other classifiers
• image normalization primitives
• Difference of Gaussian ROI detection
• Genetic Algorithm
• Phase Correlation
• Eye Detection (ASEF and MOSSE Filters)
• PCA and LDA analysis
• PCA Face Recognition
• Image Affine Transform
• Video Stream Processing

Pyvision also has a set of built in analysis tools (based on scipy) for analyzing and testing vision algorithms. This includes a simple image annotation frame work formal statistical analysis that produces confidence limits and p-values, classes designed for face detection and face recognition tests, and a table class which allows experiment data to be accumulated and then exported to csv, txt, or tex formats. PyVision also has a simple GUI (based on opencv and wxpython) which allows for the easy creation of live demos.

My goals for releasing this library is to provide this toolkit to other computer vision researchers. I would also like to have a place where other researchers contribute their code, and I could use some help making the toolkit multi-platform. The current version is focused on face recognition, and I would like some help expanding the library to other areas of computer vision and machine learning.

This library is a set of useful python classes for quickly constructing and analyze computer vision systems. The library provides a set of easy to uses object and algorithms that can be used to quickly prototype new vision algorithms.

This library contains vision algorithms that should be of general use to many people. As best as possible I will include training configurations, sample code, and unit tests to aid others in reusing the code.

My current research is focused on face recognition so much of the active development will be related that area. Contributions in other areas of research are welcome including: Object Recognition, Biological Vision, Multiview Geometry, etc.

Dependencies

• PIL
• scipy
• numpy
• scikit learn (As of pyvision 2.0)
• scikit image (As of pyvision 2.0)

Recommended

• OpenCV (optional)

Features:

• Python based for rapid prototyping.
• Common data types with easy conversion.
• Automatically tuned or easy to configure algorithms.
• Numerical support from scipy.
• Image support from PIL.
• Pretrained algorithms for common tasks.
• Python and R based algorithms analysis.
• Unit tests for algorithm verification.
• Sample code and tutorials.
• Save and load configurations with python's pickle.
• Simple interfaces to third party software: OpenCV and libsvm.

** Notes: ** Image.asOpenCV() - Requires fix to opencv bug #1618474 included in cvs on Jan 25, 2007 18:19:21 UTC by rstanchak

Releases:

• 1.3.0 - OpenCV is no longer required. This makes setup easier. Better support for "pip install pyvision_toolkit".
• 1.2.0 - Support for Python 3 and OpenCV 3
• 1.1.0 - It has been a while since the last release. OpenCV has improved recently and has become more stable with improved python bindings.
  As a result PyVision is adapting to use OpenCV for much of its core functionality. In general OpenCV tends to be faster and more accurate for common image processing functions than scipy or PIL. Additionally OpenCV will be used in the future for machine learning algorithms. The Video Task Manager (VTM) framework is improving for quickly developing video processing pipelines.
• 1.0.0 - Many interface things have been cleaned up. Function arguments that are keywords have been change and may not work as in previous version. These include: buffer, type, filter, iter. All warnings and errors from the code analysis have been fixed.
• 0.8.1 - Released in July 2010. This is the last release before a major upgrade and the last version to support OpenCV 1.0 with the swig bindings.
• 0.8.0 - Released in June 2009. This was primarily to support FaceL.