Flexible and extensible Framework for Person Re-identification problem.

PyReID allows configuring multiple preprocessing as a pipeline, followed by a Feature Extraction process and finally Feature Matching. It can calculate statistics like CMC or AUC. Not only that, but it also allows PostRanking optimization, with a functional GUI made with QT, and some embedded methods.

• Allows multiple preprocessing methods, including BTF, Illumination Normalization, Foreground/Background Segmentation using GrabCut, Symmetry-based Silhouette Partition, Static Vertical Partition and Weight maps for Weighted Histograms using Gaussian Kernels or Gaussian Mixture Models.

• Feature Extraction based on Histograms calculations. Admit 1D and 3D histograms, independent bin size for each channel, admits Regions for calculating independent Histograms per Region and Weight maps for Weighted Histograms.

• Feature Matching admits Histograms comparison methods: Correlation, Chi-Square, Intersection, Bhattacharyya distance and Euclidean distance.

• Automatically creates Ranking Matrix. For each element of the Probe obtains all Gallery elements sorted.

• Statistics Module. With the Ranking Matrix and the Dataset, obtain stats as CMC, AUC, range-X and mean value.

• Use of multiprocessing to improve speed. Preprocessing, Feature Extraction and Feature Matching are designed using multiprocessing to dramatically improve speed in multicore processors.

• Save stats in a complete excel for further review or plots creation. Save all your executions statistics in a file for later use (using Python Shelves).

Right now the project is not prepared for a direct installation as a library. In the meantime, you can check the latest sources with the command:

git clone https://github.com/Luigolas/PyReid.git

PyReID is tested to work with Python 2.7+ and the next dependencies are needed to make it work:

• OpenCV 3.0: This code makes use of OpenCV 3.0. It won't work with previous versions. This is due to incompatibility with 3D histograms and predefined values names.
• Numpy: Tested with version 1.9.2
• matplotlib: Tested with version 1.4.2
• pandas: Tested with version 0.15.0
• scikit-learn: Tested with version 0.16.1
• scipy: Tested with version 0.14.1
• xlwt: Tested with version 0.7.5

Under Construction...

This section is under construction. It is planned to have a dedicated page to further explain is usage and full potential.

Next code shows how to prepare a simple execution:

from package.dataset import Dataset
import package.preprocessing as preprocessing
from package.image import CS_HSV, CS_YCrCb
import package.feature_extractor as feature_extractor
import package.feature_matcher as feature_matcher
from package.execution import Execution
from package.statistics import Statistics

# Preprocessing
IluNormY = preprocessing.Illumination_Normalization(color_space=CS_YCrCb)
mask_source = "../resources/masks/ViperOptimalMask-4.txt"
grabcut = preprocessing.Grabcut(mask_source)  # Segmenter
preproc = [IluNormY, grabcut]  # Executes on this order

fe = feature_extractor.Histogram(CS_HSV, [16, 16, 4], "1D")
f_match = feature_matcher.HistogramsCompare(feature_matcher.HISTCMP_BHATTACHARYYA)

probe = "../datasets/viper/cam_a"
gallery = "../datasets/viper/cam_b"
ex = Execution(Dataset(probe, gallery), preproc, fe, f_match)
ranking_matrix = ex.run()

statistic = Statistics()
statistic.run(ex.dataset, ranking_matrix)

print "Range 20: %f" % statistic.CMC[19]
print "AUC: %f" % statistic.AUC

To try the GUI for PostRanking just run main.py.

Note: You will need to have a Person Re-identification Dataset to play with. You can download Viper dataset from here. Example mask seeds for GrabCut are available at resource folder.

This project is initiated as a Final Project in the University of Las Palmas by Luis González Medina.

Tutors of this project are Modesto Castrillón Santana and Javier Lorenzo Navarro

Copyright (c) 2015 Luis María González Medina.

PyReID is free software made available under the MIT License. For details see the LICENSE file.