Improved Deep Learning Architecture for Person Re-Identification

This repository implements the network described in An Improved Deep Learning Architecture for Person Re-Identification by Ahmed et al. The main deep learning library used to do this is the dlib machine learning library.

Installation

This code has only been built and tested on Ubuntu 16.04.

• dlib v19.0+
  • Requirements:
    • C++11-compatible compiler
    • CUDA 7.5 or greater
    • cuDNN v5 or greater
• CMake v2.8.12+
• HDF5 v1.8.16+
  • Used for loading the CUHK03 dataset from a MATLAB mat file.

The build is managed using CMake. In order to build this code, open a terminal and enter the following commands.

cd $THIS_REPOSITORY

mkdir build
cd build

cmake .. -DDLIB_DIR=$PATH_TO_DLIB -DGPU_ARCHITECTURE=sm_30
# ccmake . # Set BUILD_TEST to ON for unit testing. dlib flags can also be set here.

make && make install

The DLIB_DIR variable informs CMake where it should look to find dlib. This must be defined as an environment variable or passed into CMake through -DDLIB_DIR.

The optional variable GPU_ARCHITECTURE specifies what compute capability the CUDA code should be built for. By default, this variable is set to sm_30, i.e. a compute capability of 3.0. This flag is only valid if dlib detects CUDA (i.e. DLIB_USE_CUDA=ON).

Details

Pre-processing

Global contrast normalization is applied to each image at the input layer.

Architecture Modifications

• Each 5x5 convolutional layer, except for the "patch summary features" layer, has been replaced by two 3x3 convolutional layers, with batch normalization after each.
• Batch normalization was added after the fully connected layer.

Training Modifications

• Minibatches consist of 128 image pairs, with an even split between positive and negative examples.
• No hard negative and data augmentation were used for training.

Results

Below is a cumulative match curve (CMC) produced by the network implemented in this repository. The criteria uses the evaluation as described here (repository for Domain Guided Dropout for Person Re-ID).

Currently, only CUHK03 training and testing has been implemented (in cuhk03.cpp).