Person Re-identification

This is a person re-identification (re-id) project that aims at learning generic deep features from multiple person re-id datasets.

Installation

We have integrated our self-brewed caffe into external/caffe, which provides batch-normalization and multi-gpu parallel training. Please clone this project with the command:

git clone --recursive https://github.com/Cysu/person_reid.git

Apart from the official installation prerequisites, we have several other dependencies: cudnn-v4, openmpi, and 1.55 <= boost < 1.60. You may install them manually or by a package manager (apt-get, pacman, yum, etc.).

Then configure the Makefile.config and compile the caffe.

cd external/caffe
cp Makefile.config.example Makefile.config
# Configure the libraries properly
make -j8 && make py

Setup environment for experiments

First, download some person re-id datasets at here with password 8hjx. Make a soft link to the root directory:

ln -sf /path/to/the/root/of/datasets external/raw

Next, create a directory for our experiments, where later we will put formatted datasets, databases, snapshots, and some results.

mkdir -p /path/to/the/experiment/directory
ln -sf /path/to/the/experiment/directory external/exp

Prepare data

First we need to format raw datasets into our uniform data format.

scripts/format_rawdata.sh

Next convert each formatted dataset into a LMDB.

scripts/make_dbs.sh

At last merge all the datasets together for the joint single-task learning (JSTL).

scripts/merge_dbs.sh

Train nets

Some model and solver definitions are listed in models/. Use the shell scripts provided in scripts/ to run the experiments. Note that by default the scripts will use two GPUs. You may find the command mpirun -n 2 ... -gpu 0,1 in the scripts and adapt it with your own settings.

As an example, training a net by Joint Single-Task Learning (JSTL) is done by

scripts/exp_jstl.sh 0 googlenet_bn

The snapshots will be saved in external/exp/snapshots/jstl/.

Evaluate the performance

After training, use scripts/extract_features.sh dataset_split_name model_name caffemodel_path [blob_name] to extract features, for example,

scripts/extract_features.sh cuhk03_split_00 googlenet_bn \
  external/exp/snapshots/jstl/jstl_split_00_googlenet_bn_iter_55000.caffemodel \
  fc7_bn

Then learn a metric and evaluate by CMC score, for example,

python2 eval/metric_learning.py \
  external/exp/results/cuhk03_split_00_googlenet_bn_fc7_bn_jstl_split_00_googlenet_bn_iter_55000

This will print several top-k accuracies, and you may find the top-1 accuracy to be around 73%.

Datasets

We summarize some commonly used person re-id datasets below. They can be downlaoded from here with password 8hjx.

Name	Reference
3DPeS	D. Baltieri, et al., 3DPes: 3D people dataset for surveillance and forensics
CUHK01	W. Li, et al., Human reidentification with transferred metric learning
CUHK02	W. Li, et al., Locally Aligned Feature Transforms across Views
CUHK03	W. Li, et al., Deepreid: Deep filter pairing neural network for person re-identification
i-LIDS	W. Zheng, et al., Associating groups of people
i-LIDS-VID	T. Wang, et al., Person Re-Identification by Video Ranking
Market-1501	L. Zheng, et al., Scalable Person Re-identification: A Benchmark
OPeRID	S. Liao, et al., Open-set Person Re-identification
PRID	M. Hirzer, et al., Person re-identification by descriptive and discriminative classification
RAiD	A. Das et al., Consistent re-identification in a camera network
Shinpuhkan	Y. Kawanishi, et al., Shinpuhkan2014: A Multi-Camera Pedestrian Dataset for Tracking People across Multiple Cameras
VIPeR	D. Gray, et al., Evaluating appearance models for recognition, reacquisition, and tracking