instance-retrieval-reamde.md

Instance Retrieval

Person Re-identification is adopted for evaluating the performance on the instance retrieval task.

Prerequisites

• Create a conda virtual environment and activate it.

conda create --name uncertainty python=3.6
conda activate uncertainty

• Install Pytorch and torchvision following the official instructions, e.g.,

conda install pytorch==1.5.0 torchvision==0.6.0 cudatoolkit=10.1 -c pytorch

• Install the dependent libraries.

pip install numpy six h5py Pillow scipy scikit-learn metric-learn

Prepare Datasets

• Download the raw datasets DukeMTMC-reid and Market-1501, then unzip them under the directory like:

instance-retrieval/examples/data
├── dukemtmc
│   └── DukeMTMC-reID
└── market1501
    └── Market-1501-v15.09.15
...

Installation

cd instance-retriveal
python setup.py develop

Getting Started

We utilize 4 Nvidia Tesla V100 (32GB) for training.

• Train models on dukemtmc and test on unseen market1501:

sh scripts/train.sh dukemtmc market1501 uresnet50 1 uncertainty

• Train models on market1501 and test on unseen dukemtmc:

sh scripts/train.sh market1501 dukemtmc uresnet50 1 uncertainty

The checkpoints and logs can be found at link.

Acknowledge

The implementation is based on MMT. We thank them for their excellent projects.