A Discriminatively Learned CNN Embedding for Person Re-identification

A caffe-based implementation of this paper, providing whole training, testing and evaluation codes.

The official code (written in matconvnet) is available here.

Preparation

• download Market-1501 dataset and put Market-1501 in $HOME/datasets/
• cd examples/market1501/mat-codes and matlab -nodisplay -r 'generate_train(), exit()' to generate train, test and qurey data lists.
• Build with NCLL / cuda-8.0 / cudnn-v5.1

Results on Market-1501

Market-1501 is one of the most popular person re-identification datasets.

Models can be found in models/market1501/model_name

Many scripts (e.g initialization, testing, training, extract feature and evaluation) can be found in examples/market1501/

[iter_size * batch_size] = real batch_size

CaffeNet

• Link to the pre-trained CaffeNet model
• python models/market1501/generate_caffenet.py for generate caffenet based person re-ID network and solver files.
• sh models/market1501/caffenet/train.sh --gpu 0 for training models.
• sh examples/market1501/extract/extract_prediction.sh 0 caffenet fc7 for extracting features of query and test data
• cd examples/market1501/evaluation/ and evaluation('caffenet') to evaluate performance of the trained model on Market-1501
• final results are [1x128] : mAP = 0.402689, r1 precision = 0.639846 [Euclidean]

GoogleNet

• Link to the pre-trained GoogleNet model
• GoogleNet-v1 model is already in models/market1501/googlenet
• sh models/market1501/googlenet/train.sh --gpu 0
• sh examples/market1501/extract/extract_prediction.sh 0 googlenet pool5/7x7_s1
• cd examples/market1501/evaluation/ and evaluation('googlenet')
• final results are : mAP = 0.511545, r1 precision = 0.735154 [Cos + Eucl]

VGG-16

• Link to the pre-trained VGG-16 model
• python models/market1501/generate_vgg16.py for generate caffenet based person re-ID network and solver files.
• sh models/market1501/vgg16/train.sh --gpu 2,3 for training
• sh examples/market1501/extract/extract_prediction.sh 0 vgg16 fc7 for extracting features
• cd examples/market1501/evaluation/ and evaluation('vgg16') to evaluate performance of vgg16/fc7 on Market-1501
• final results are [2x 24] : mAP = 0.456417, r1 precision = 0.677257

resnet-50

• python models/market1501/generate_resnet50.py
• sh models/market1501/res50/train.sh --gpu 2,3
• sh examples/market1501/extract/extract_prediction.sh 0 res50 pool5
• final results are : mAP = 0.585765, r1 precision = 0.790974 [Cos + Eucl]

Citation

Please cite the following papers in your publications if it helps your research:

@article{zheng2016discriminatively,
  title={A Discriminatively Learned CNN Embedding for Person Re-identification},
  author={Zheng, Zhedong and Zheng, Liang and Yang, Yi},
  journal={TOMM},
  year={2017}
}
@article{jia2014caffe,
  Author = {Jia, Yangqing and Shelhamer, Evan and Donahue, Jeff and Karayev, Sergey and Long, Jonathan and Girshick, Ross and Guadarrama, Sergio and Darrell, Trevor},
  Journal = {arXiv preprint arXiv:1408.5093},
  Title = {Caffe: Convolutional Architecture for Fast Feature Embedding},
  Year = {2014}
}