GLAD
This repository provides the code for our ACM MM17 paper GLAD: Global-Local-Alignment Descriptor for Pedestrian Retrieval
Step.1 Pose Estimation
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.
Step.2 Descriptor Learning
Make our caffe
We have modify the original caffe, please make our provided caffe before running our code.
Dataset
Download Market1501 Dataset. Then process these raw data as step.1.
ImageNet Pretrained model
Download GoogLeNet model pretrained on Imagenet.
Train our GLAD
- Modify the
prototxt\train_val.prototxt
. Please modify the dataset path in the file. - End up training with 10,0000 iterations. More details, please see the
prototxt\solver_stepsize_6400_2_step3_ver4_65.prototxt
Step.3 Test
- Extract fc6(and layer1/fc6, layer2/fc6, layer3/fc6) features.
- L1 normalization is needed.
- Adding weights for these four features according to our paper.
Our Model
- If you require our trained model, please contact Longhui Wei(weilh2568@gmail.com).
- If you have any questions about our code or paper, please contact Longhui Wei
Citation
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}
}