Domain Guided Dropout for Person Re-identification
Switch branches/tags
Nothing to show
Clone or download
Latest commit 19c2a05 Dec 20, 2017

README.md

Domain Guided Dropout for Person Re-id

This project aims at learning generic person re-identification (re-id) deep features from multiple datasets with domain guided dropout. Mainly based on our CVPR 2016 paper Learning Deep Feature Representations with Domain Guided Dropout for Person Re-identification.

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/dgd_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 (a tip for installing boost 1.55 on Ubuntu 14.04: sudo apt-get autoremove libboost1.54* then sudo apt-get install libboost1.55-all-dev).

Then configure the Makefile.config and compile the caffe. To use multi-GPU for training, please uncomment the MPI parallel block in the Makefile.config and set the MPI_INCLUDE and MPI_LIB properly. Please find more details of using the caffe here.

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

Some other prerequisites are

  1. Matlab (to pre-process the CUHK03 dataset)
  2. python2 packages: numpy, scipy, Pillow, scikit-learn, protobuf, lmdb
  3. Add export PYTHONPATH=".:$PYTHONPATH" to ~/.bashrc and restart the terminal

Download datasets

Download the following datasets.

  1. CUHK03
  2. CUHK01
  3. PRID
  4. VIPeR (I cannot find the link to the original dataset. This is my previous backup version.)
  5. 3DPeS
  6. i-LIDS (I cannot find the link to the original dataset. This is my previous backup version.)
  7. Shinpuhkan (need to send an email to the authors)

Link the root directory of these datasets to our project.

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

Prepare data

  1. Create a directory for experiment data and results

    mkdir -p external/exp
    

    or link against another external directory

    ln -s /path/to/your/exp/directory external/exp
    
  2. Convert raw datasets into a uniform data format

    scripts/format_rawdata.sh
    
  3. Convert formatted datasets into LMDBs

    scripts/make_dbs.sh
    
  4. Merge all the datasets together for the joint single-task learning (JSTL)

    scripts/merge_dbs.sh
    

Experiments

Note: We use two GPUs to train the models by default. Change the mpirun -n 2 ... -gpu 0,1 in scripts/routines.sh to your own hardware configuration if necessary.

**GPU device id needs to be changed in:

  1. train_model() and extract_features() of scripts/rountines.sh
  2. main() of tools/compute_impact_score.py**

Our experiments are organized into two groups:

  1. Baseline: training individually on each dataset
  2. Ours: Joint single task learning (JSTL) + Domain guided dropout (DGD)

We provide a pretrained JSTL+DGD model here that can be used as a generic person re-id feature extractor.

Some archived experiment logs can be found at archived/.

Baseline: training individually on each dataset

To train and test a model individually on a dataset, just run the script

scripts/exp_individually.sh prid

where the parameter is the dataset name, can be one of cuhk03, cuhk01, prid, viper, 3dpes, ilids.

Ours: Joint single task learning (JSTL) + Domain guided dropout (DGD)

  1. Pretrain a model using the mixed dataset with JSTL. The CMC accuracies printed out are corresponding to the JSTL entries in Table 3 of our paper.

     scripts/exp_jstl.sh
    
  2. Based on the pretrained JSTL model, we first compute the neuron impact scores (NIS) for each dataset, and then resume the JSTL training with deterministic DGD. The CMC accuracies printed out are corresponding to the JSTL+DGD entries in Table 3 of our paper.

     scripts/exp_dgd.sh
    

    At last, to achieve the best performance, we can fine-tune the model on each dataset with stochastic DGD. The CMC accuracies printed out are corresponding to the FT-(JSTL+DGD) entries in Table 3 of our paper.

     scripts/exp_ft_dgd.sh
    

Citation

@inproceedings{xiao2016learning,
  title={Learning Deep Feature Representations with Domain Guided Dropout for Person Re-identification},
  author={Xiao, Tong and Li, Hongsheng and Ouyang, Wanli and Wang, Xiaogang},
  booktitle={CVPR},
  year={2016}
}

Referenced Datasets

We summarize the person re-id datasets used in this project as below.

Name Reference
3DPeS Baltieri, et al., 3DPes: 3D people dataset for surveillance and forensics
CUHK01 Li, et al., Human reidentification with transferred metric learning
CUHK02 Li, et al., Locally Aligned Feature Transforms across Views
CUHK03 Li, et al., Deepreid: Deep filter pairing neural network for person re-identification
i-LIDS Zheng, et al., Associating groups of people
PRID Hirzer, et al., Person re-identification by descriptive and discriminative classification
Shinpuhkan Kawanishi, et al., Shinpuhkan2014: A Multi-Camera Pedestrian Dataset for Tracking People across Multiple Cameras
VIPeR Gray, et al., Evaluating appearance models for recognition, reacquisition, and tracking