futuretwt / hmah Goto Github PK

Our training environment(only as reference):

Python 3.6.2
Pytorch 1.5.0
RTX2080Ti + CUDA 10.2

You can easily train the model of our method with the following code:

chmod a+x ./train.sh
./train.sh

(1) Firstly, you need to divide the dataset into three parts: train, query, retrieval. Each part should contain image features, text features and labels.

(2) Then, place the partitioned dataset in the '/Data' directory.

(3) Finally, change the corresponding parameters and train the model.

You can easily reimplement the experimental results using the above code. The dataset link is provided in Section.4. If the results you reimplement differ from the below results, it may be due to machine differences. The performances on MS COCO are shown as follow, and the other results can refer our paper.

[16 bits]
    mAP@500 on I2T: 0.691
    mAP@500 on T2I: 0.800

[32bits]
    mAP@500 on I2T: 0.732
    mAP@500 on T2I: 0.869

[64bits]
    mAP@500 on I2T: 0.763
    mAP@500 on T2I: 0.904

[128bits]
    mAP@500 on I2T: 0.790
    mAP@500 on T2I: 0.934

[256bits]
    mAP@500 on I2T: 0.808
    mAP@500 on T2I: 0.947

Google Drive Link:
    https://drive.google.com/drive/folders/1-Ru-NOaIukbJj_wL1rUh7jThOTGwPmKM?usp=sharing

If you want to run our code compared with all the cross-modal hashing retrieval baselines on three datasets, we suggest that you should refer the follow link. In this repository, all the datasets are provided and all the baselines can be run easily. You only need change the data path dict "paths_#" and the loaded data key of funtion "load_#()" in "utils/datasets.py".

https://github.com/BMC-SDNU/Cross-Modal-Hashing-Retrieval

@ARTICLE{HMAH,  
    author={Tan, Wentao and Zhu, Lei and Li, Jingjing and Zhang, Huaxiang and Han, Junwei},  
    journal={IEEE Transactions on Multimedia},   
    title={Teacher-Student Learning: Efficient Hierarchical Message Aggregation Hashing for Cross-Modal Retrieval},   
    year={2022},  
    pages={1-1},  
    doi={10.1109/TMM.2022.3177901}
}