This repository provides the code for our ACM MM17 paper GLAD: Global-Local-Alignment Descriptor for Pedestrian Retrieval

The first stage is to estimate the human keypoints. We used the deepercut model provided in DeeperCut. Especially, we utilize the single person pose estimation model provided by the authors.

Afer pose estimation, please detect the three parts according to our paper. Example image is as followes:

You can utilize any pose estimation methods to replace DeeperCut.

We have modify the original caffe, please make our provided caffe before running our code.

Download Market1501 Dataset. Then process these raw data as step.1.

Download GoogLeNet model pretrained on Imagenet.

1. Modify the prototxt\train_val.prototxt. Please modify the dataset path in the file.
2. End up training with 10,0000 iterations. More details, please see the prototxt\solver_stepsize_6400_2_step3_ver4_65.prototxt

1. Extract fc6(and layer1/fc6, layer2/fc6, layer3/fc6) features.
2. L1 normalization is needed.
3. Adding weights for these four features according to our paper.

1. If you require our trained model, please contact Longhui Wei([email protected]).
2. If you have any questions about our code or paper, please contact Longhui Wei

Please cite our paper in your publications if it helps your research:

@inproceedings{wei2017glad,
  title={GLAD: Global-Local-Alignment Descriptor for Pedestrian Retrieval},
  author={Wei, Longhui and Zhang, Shiliang and Yao, Hantao and Gao, Wen and Tian, Qi},
  booktitle={ACM MM},
  year={2017}
}