Awesome

Improving One-stage Visual Grounding by Recursive Sub-query Construction

by Zhengyuan Yang, Tianlang Chen, Liwei Wang, and Jiebo Luo

European Conference on Computer Vision (ECCV), 2020

Introduction

We propose a recursive sub-query construction framework to address previous one-stage visual grounding methods' limitations on grounding long and complex queries. For more details, please refer to our paper.

[1] Yang, Zhengyuan, et al. "A fast and accurate one-stage approach to visual grounding". ICCV 2019.

Prerequisites

Python 3.6 (3.5 tested)
Pytorch 0.4.1 and 1.4.0 tested (other versions in between should work)
Others (Pytorch-Bert, etc.) Check requirements.txt for reference.

Installation

Clone the repository

git clone https://github.com/zyang-ur/ReSC.git

Prepare the submodules and associated data

RefCOCO, RefCOCO+, RefCOCOg, ReferItGame Dataset: place the data or the soft link of dataset folder under ./ln_data/. We follow dataset structure DMS. To accomplish this, the download_dataset.sh bash script from DMS can be used.
```
bash ln_data/download_data.sh --path ./ln_data
```

Data index: download the generated index files and place them as the ./data folder. Availble at [Gdrive], [One Drive].
```
rm -r data
tar xf data.tar
```
Model weights: download the pretrained model of Yolov3 and place the file in ./saved_models.
```
sh saved_models/yolov3_weights.sh
```

More pretrained models are availble in the performance table [Gdrive], [One Drive] and should also be placed in ./saved_models.

Training

Train the model, run the code under main folder. Using flag --large to access the ReSC-large model. ReSC-base is the default.

python train.py --data_root ./ln_data/ --dataset referit \
  --gpu gpu_id --resume saved_models/ReSC_base_referit.pth.tar

Evaluate the model, run the code under main folder. Using flag --test to access test mode.

python train.py --data_root ./ln_data/ --dataset referit \
  --gpu gpu_id --resume saved_models/ReSC_base_referit.pth.tar --test

Implementation Details

We train 100 epoches with batch size 8 on all datasets expect RefCOCOg, where we find training 20/40 epoches have the best performance. We fix the bert weights during training as the default. The language encoder can be finetuned with the flag --tunebert. We observe a small improvenment on some datasets (e.g. RefCOCOg). Please check other experiment settings in our paper.

Performance and Pre-trained Models

Pre-trained models are availble in [Gdrive], [One Drive].

<table> <thead> <tr> <th>Dataset</th> <th>Ours-base (Acc@0.5)</th> <th>Ours-large (Acc@0.5)</th> </tr> </thead> <tbody> <tr> <td rowspan=3>RefCOCO</td> <td>val: 76.74</td> <td>val: 78.09</td> </tr> <tr> <td>testA: 78.61</td> <td>testA: 80.89</td> </tr> <tr> <td>testB: 71.85</td> <td>testB: 72.97</td> </tr> <tr> <td rowspan=3>RefCOCO+</td> <td>val: 63.21</td> <td>val: 62.97</td> </tr> <tr> <td>testA: 65.94</td> <td>testA: 67.13</td> </tr> <tr> <td>testB: 56.08</td> <td>testB: 55.43</td> </tr> <tr> <td rowspan=3>RefCOCOg</td> <td>val-g: 61.12</td> <td>val-g: 62.22</td> </tr> <tr> <td>val-umd: 64.89</td> <td>val-umd: 67.50</td> </tr> <tr> <td>test-umd: 64.01</td> <td>test-umd: 66.55</td> </tr> <tr> <td rowspan=2>ReferItGame</td> <td>val: 66.78</td> <td>val: 67.15</td> </tr> <tr> <td>test: 64.33</td> <td>test: 64.70</td> </tr> </tbody> </table>

Citation

@inproceedings{yang2020improving,
  title={Improving One-stage Visual Grounding by Recursive Sub-query Construction},
  author={Yang, Zhengyuan and Chen, Tianlang and Wang, Liwei and Luo, Jiebo},
  booktitle={ECCV},
  year={2020}
}
@inproceedings{yang2019fast,
  title={A Fast and Accurate One-Stage Approach to Visual Grounding},
  author={Yang, Zhengyuan and Gong, Boqing and Wang, Liwei and Huang
    , Wenbing and Yu, Dong and Luo, Jiebo},
  booktitle={ICCV},
  year={2019}
}

Credits

Our code is built on Onestage-VG.

Part of the code or models are from DMS, film, MAttNet, Yolov3 and Pytorch-yolov3.