ZhangLeUestc / caffe-reid

A Discriminatively Learned CNN Embedding for Person Re-identification

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.

Data Preparation

• download Market-1501 dataset and ln -s $Market-1501 examples/market1501/
• cd examples/market1501/mat-codes and run generate_train.m to generate train, test and qurey data lists.

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

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

googlenet

• GoogleNet-v1 model is already in models/market1501/googlenet
• training and testing processes are similar as previous
• final results are [2x 64] : mAP = 0.476404, r1 precision = 0.706948 [Euclidean]
• final results are [2x 64] : mAP = 0.489998, r1 precision = 0.710214 [Cos + Eucl]

vgg16

• python models/market1501/generate_vgg16.py for generate caffenet based person re-ID network and solver files.
• training and testing processes are similar as previous
• final results are [2x 36] : mAP = 0.446430, r1 precision = 0.654394 [Failed]

vgg-reduce

the atrous version of VGG16 (Semantic image segmentation with deep convolutional nets and fully connected crfs)

• final results are [2x 32] : mAP = 0.461156, r1 precision = 0.719715 [Cos + Eucl] [Global MAX Pooling]
• By dropping the global max pooling layer in training and testing phase, a better performance can be obtained
• final results are [2x 32] : mAP = 0.511268, r1 precision = 0.745843 [Cos + Eucl] [No Pooling]

resnet-50

• final results are [4x 16] : mAP = 0.593053, r1 precision = 0.801960 [Cos + Eucl]

Citation

Please cite this paper 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={arXiv preprint arXiv:1611.05666},
  year={2016}
}

Caffe

Caffe is a deep learning framework made with expression, speed, and modularity in mind. It is developed by the Berkeley Vision and Learning Center (BVLC) and community contributors.

Check out the project site for all the details like

• DIY Deep Learning for Vision with Caffe
• Tutorial Documentation
• BVLC reference models and the community model zoo
• Installation instructions

and step-by-step examples.

Please join the caffe-users group or gitter chat to ask questions and talk about methods and models. Framework development discussions and thorough bug reports are collected on Issues.

Happy brewing!

License and Citation

Caffe is released under the BSD 2-Clause license. The BVLC reference models are released for unrestricted use.

Please cite Caffe in your publications if it helps your research:

@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}
}