Introduction

Code for our AAAI 2021 paper Confidence-aware Non-repetitive Multimodal Transformers for TextCaps [PDF].

Installation

Our implementation is based on Pythia framework (now called mmf), and built upon M4C-Captioner. Please refer to Pythia's document for more details on installation requirements.

# install pythia based on requirements.txt
python setup.py build develop  

Data Preparation

The following is open-source data of TextCaps dataset from M4C-Captioner's Github repository. Please download them from the links below and and extract them under data directory.

• object Faster R-CNN Features of TextCaps
• OCR Faster R-CNN Features of TextCaps
• detectron weights of TextCaps

Our imdb files include new OCR tokens and recognition confidence extracted with pretrained OCR systems ( CRAFT, ABCNet and four-stage STR). The three imdb files should be downloaded from the links below and put under data/imdb/.

Finally, your data directory structure should look like this:

data
|-detectron                         
|---...
|-m4c_textvqa_ocr_en_frcn_features
|---...
|-open_images                       
|---...
|-vocab_textcap_threshold_10.txt    #already provided
|-imdb                              
|---imdb_train.npy                  
|---imdb_val_filtered_by_image_id.npy   
|---imdb_test_filtered_by_image_id.npy      

Pretrained Model

Training

We provide an example script for training on TextCaps dataset for 12000 iterations and evaluating every 500 iterations.

./train.sh

This may take approximately 13 hours, depending on GPU devices. Please refer to our paper for implementation details.

First-time training will download fasttext model . You may also download it manually and put it under pythia/.vector_cache/.

During training, log file can be found under save/cnmt/m4c_textcaps_cnmt/logs/. You may also run training in background and check log file for training status.

Evaluation

Assume that checkpoint of the trained model is saved at save/cnmt/m4c_textcaps_cnmt/best.ckpt (otherwise modify the resume_file parameter in the shell script).

Run the following script to generate prediction json file:

#evaluate on validation set
./eval_val.sh 
#evaluate on test set
./eval_test.sh

The prediction json file will be saved under save/eval/m4c_textcaps_cnmt/reports/. You can submit the json file to the TextCaps EvalAI server for result.

Citation

@article{wang2020confidenceaware,
  title={Confidence-aware Non-repetitive Multimodal Transformers for TextCaps}, 
  author={Wang, Zhaokai and Bao, Renda and Wu, Qi and Liu, Si},
  year={2020},
  journal={arXiv preprint arXiv:2012.03662},
}