/FGCrossNet_ACMMM2019

Source code of our ACM MM 2019 paper "A New Benchmark and Approach for Fine-grained Cross-media Retrieval".

Primary LanguagePython

A New Benchmark and Approach for Fine-grained Cross-media Retrieval

Introduction

This repository contains the pytorch codes, trained models, and the new benchmarks described in our ACM MM 2019 paper "A New Benchmark and Approach for Fine-grained Cross-media Retrieval".

For more details, please visit our project page.

Results

  • The MAP scores of bi-modality fine-grained cross-media retrieval of our FGCrossNet
I->T I->A I->V T->I T->A T->V A->I A->T A->V V->I V->T V->A Average
FGCrossNet(ours) 0.210 0.526 0.606 0.255 0.181 0.208 0.553 0.159 0.443 0.629 0.195 0.437 0.366
  • The MAP scores of multi-modality fine-grained cross-media retrieval of our FGCrossNet
I->All T->All V->All A->All Average
FGCrossNet(ours) 0.549 0.196 0.416 0.485 0.412

Installation

Requirement

  • pytorch, tested on [v1.0]
  • CUDA, tested on v9.0
  • Language: Python 3.6

1. Download dataset

Please visit our project page.

2. Download trained model

The trained models of our FGCrossNet framework can be downloaded from OneDrive, Google Drive or Baidu Cloud.

3. Prepare audio data

python audio.py

4. Training

sh train.sh

5. Testing

sh test.sh

Citing

@inproceedings{he2019fine,
    Author = {Xiangteng He, Yuxin Peng, Liu Xie},
    Title = {A New Benchmark and Approach for Fine-grained Cross-media Retrieval},
    Booktitle = {Proc. of ACM International Conference on Multimedia (ACM MM)},
    Year = {2019}
} 

Contributing

For any questions, feel free to open an issue or contact us. (hexiangteng@pku.edu.cn)