person re-identification

Simple python model for person re-identification. Use as follows:

import cv2
from reid import reid

im1 = cv2.cvtColor(cv2.imread('/path/to/file1.png'), cv2.COLOR_BGR2RGB)
im2 = cv2.cvtColor(cv2.imread('/path/to/file2.png'), cv2.COLOR_BGR2RGB)

# im1 / im2 can either be images or lists of images. In case of lists
# prediction is done item-wise

model = reid.ReId()
score = model.predict(im1, im2)

The model will give a score below 0.5 if it believes the two persons are not the same and a value larger than 0.5 if it thinks it got the same people. Below are some sample pairs of images the model has never seen before (thx Joao):

Install

This library was developed using Python 3.6, numpy, Keras 2.1.3. and tensorflow 1.8 so make sure you have them installed. Furthermore, to manage the datasets and to provide some utilities I use pak which you can install as follows:

pip install git+https://github.com/jutanke/pak.git

Last but not least, OpenCV3 is used for image manipulation etc. so you should install that one as well. Usually, I compile it from source like so:

git clone https://github.com/opencv/opencv.git
cd opencv && git checkout 3.4.0
mkdir build && cd build
cmake -DBUILD_opencv_java=OFF \ 
    -DBUILD_TESTS=OFF -DBUILD_PERF_TESTS=OFF -DCMAKE_BUILD_TYPE=RELEASE \
    -DCMAKE_INSTALL_PREFIX=$(python3 -c "import sys; print(sys.prefix)") \
    -DPYTHON3_EXECUTABLE=$(which python3) \
    -DPYTHON3_INCLUDE_DIR=$(python3 -c "from distutils.sysconfig import get_python_inc; print(get_python_inc())") \
    -DPYTHON3_PACKAGES_PATH=$(python3 -c "from distutils.sysconfig import get_python_lib; print(get_python_lib())") .. 
make -j4
make install

but there seems to be a package for anaconda that you might install like:

conda install -c menpo opencv3 

(I have not tried this one though).

After you have covered all prerequisite you can install this library as follows:

pip install git+https://github.com/jutanke/person_reid.git

Model

The model is a quite simple Siamese Network using DenseNet as feature extractor.

Below the training and validation accuracy is being reported: