This repository contains the reference code for our paper RSTNet: Captioning with Adaptive Attention on Visual and Non-Visual Words (CVPR 2021).

If you have any questions about our work, feel free to post issues on this github project. I will answer your questions and update code monthly.
If you are in hurry, please email me via [email protected].
If our work is helpful to you or gives some inspiration to you, please star this project and cite our paper. Thank you!

@inproceedings{zhang2021rstnet,
  title={RSTNet: Captioning with adaptive attention on visual and non-visual words},
  author={Zhang, Xuying and Sun, Xiaoshuai and Luo, Yunpeng and Ji, Jiayi and Zhou, Yiyi and Wu, Yongjian and Huang, Feiyue and Ji, Rongrong},
  booktitle={Proceedings of the IEEE/CVF conference on computer vision and pattern recognition},
  pages={15465--15474},
  year={2021}
}

Clone the repository and create the m2release conda environment using the environment.yml file:

conda env create -f environment.yml
conda activate m2release

Then, download spacy data by executing the following command:
python -m spacy download en or python -m spacy download en_core_web_sm.

You also need to create 5 new folders, namely Datasets, save_language_models, language_tensorboard_logs, save_transformer_models and transformer_tensorboard_logs in the root directory of this project.

To run our code, you need to put annotations folder m2_annotations, visual features folder X101-features for the COCO dataset into Datasets.

Most annotations have been prepared by [1], please download m2_annotations and put it into this root directory.

Visual features are computed with the code provided by [2]. To reproduce our result, please download the COCO features file such as X-101-features.tgz in grid-feats-vqa and rename the extracted folder as X101-features. Considering that this feature file is huge, you can alternatively save the features as float16 for storage space saving by executing the following command:

python switch_datatype.py

In order to solve the shape difference and match the feat shape with region feat shape (50 regions), please execute the following command to reshape the visual to 49(7x7) and save all visual features as a h5py file.

python feats_process.py

Note that you can also access to my processed offline image features coco_grid_feats in Baidu Netdisk with the extraction code cvpr for convenience.

In addition, if you want to extract the grid-based features of your custom image dataset, you can refer to the codes in project grid-feats-vqa .

Run python train_language.py and python train_transformer.py in sequence using the following arguments:

For example, to train our BERT-based language model with the parameters used in our experiments, use

python train_language.py --exp_name bert_language --batch_size 50 --features_path /path/to/features --annotation_folder /path/to/annotations

to train our rstnet model with the parameters used in our experiments, use

python train_transformer.py --exp_name rstnet --batch_size 50 --m 40 --head 8 --features_path /path/to/features --annotation_folder /path/to/annotations

The figure below shows the changes of cider value during the training of rstnet. You can also visualize the training details by calling the tensorboard files in tensorboard_logs.

Run python test_offline.py to evaluate the performance of rstnet on the Karpathy test split of MS COCO dataset.

Run python test_online.py to generate required files and evaluate the performance of rstnet on the official test server of MS COCO dataset.

Note that, to reproduce the our reported results, you can also download our pretrained model files in the pretrained_models folder and put them into folder saved_language_models and folder saved_language_models repectively . The results of offline evaluation (Karpathy test split of MS COCO) are as follows：

[1] Cornia, M., Stefanini, M., Baraldi, L., & Cucchiara, R. (2020). Meshed-memory transformer for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.
[2] Jiang, H., Misra, I., Rohrbach, M., Learned-Miller, E., & Chen, X. (2020). In defense of grid features for visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.

Thank Cornia et.al for their open source code meshed-memory-transformer , on which our implements are based.
Thank Jiang et.al for the significant discovery in visual representation [2], which has given us a lot of inspiration.