Home

Awesome

Meta Self-Learning for Multi-Source Domain Adaptation: A Benchmark Tweet

Project | Arxiv | YouTube | PWC | PWC

News


dataset1

Abstract

In recent years, deep learning-based methods have shown promising results in computer vision area. However, a common deep learning model requires a large amount of labeled data, which is labor-intensive to collect and label. What’s more, the model can be ruined due to the domain shift between training data and testing data. Text recognition is a broadly studied field in computer vision and suffers from the same problems noted above due to the diversity of fonts and complicated backgrounds. In this paper, we focus on the text recognition problem and mainly make three contributions toward these problems. First, we collect a multi-source domain adaptation dataset for text recognition, including five different domains with over five million images, which is the first multi-domain text recognition dataset to our best knowledge. Secondly, we propose a new method called Meta Self-Learning, which combines the self-learning method with the meta-learning paradigm and achieves a better recognition result under the scene of multi domain adaptation. Thirdly, extensive experiments are conducted on the dataset to provide a benchmark and also show the effectiveness of our method.

Data Prepare

Download the dataset from here.

Before using the raw data, you need to convert it to lmdb dataset.

python create_lmdb_dataset.py --inputPath data/ --gtFile data/gt.txt --outputPath result/

The data folder should be organized below

data
├── train_label.txt
└── imgs
    ├── 000000001.png
    ├── 000000002.png
    ├── 000000003.png
    └── ...

The format of train_label.txt should be {imagepath}\t{label}\n For example,

imgs/000000001.png Tiredness
imgs/000000002.png kills
imgs/000000003.png A

Requirements

Argument

Get started

git clone https://github.com/bupt-ai-cz/Meta-SelfLearning.git
cd Meta-SelfLearning

To train the baseline model for synthetic domain.

OMP_NUM_THREADS=8 CUDA_VISIBLE_DEVICES=0 python train.py \
    --train_data data/train/ \
    --select_data car-doc-street-handwritten \
    --batch_ratio 0.25-0.25-0.25-0.25 \
    --valid_data data/test/syn \
    --Transformation None --FeatureExtraction ResNet \
    --SequenceModeling BiLSTM --Prediction Attn \
    --batch_size 96 --valInterval 5000

To train the meta_train model for synthetic domain using the pretrained model.

OMP_NUM_THREADS=8 CUDA_VISIBLE_DEVICES=0 python meta_train.py 
    --train_data data/train/ \ 
    --select_data car-doc-street-handwritten \
    --batch_ratio 0.25-0.25-0.25-0.25 \
    --valid_data data/test/syn/ \
    --Transformation None --FeatureExtraction ResNet \
    --SequenceModeling BiLSTM --Prediction Attn \
    --batch_size 96  --source_num 4  \
    --valInterval 5000 --inner_loop 1\
    --saved_model saved_models/pretrained.pth 

To train the pseudo-label model for synthetic domain.

OMP_NUM_THREADS=8 CUDA_VISIBLE_DEVICES=0 python self_training.py 
    --train_data data/train \
    —-select_data car-doc-street-handwritten \
    --batch_ratio 0.25-0.25-0.25-0.25 \
    --valid_data data/train/syn \
    --test_data data/test/syn \
    --Transformation None --FeatureExtraction ResNet \
    --SequenceModeling BiLSTM --Prediction Attn \
    --batch_size 96  --source_num 4 \
    --warmup_threshold 28 --pseudo_threshold 0.9 \
    --pseudo_dataset_num 50000 --valInterval 5000 \ 
    --saved_model saved_models/pretrained.pth 

To train the meta self-learning model for synthetic domain.

OMP_NUM_THREADS=8 CUDA_VISIBLE_DEVICES=0 python meta_self_learning.py 
    --train_data data/train \
    —-select_data car-doc-street-handwritten \
    --batch_ratio 0.25-0.25-0.25-0.25 \
    --valid_data data/train/syn \
    --test_data data/test/syn \
    --Transformation None --FeatureExtraction ResNet \
    --SequenceModeling BiLSTM --Prediction Attn \
    --batch_size 96 --source_num 4 \
    --warmup_threshold 0 --pseudo_threshold 0.9 \
    --pseudo_dataset_num 50000 --valInterval 5000 --inner_loop 1 \
    --saved_model pretrained_model/pretrained.pth 

Citation

If you use this data or code for your research, please cite our paper Meta Self-Learning for Multi-Source Domain Adaptation: A Benchmark

@inproceedings{qiu2021meta,
  title={Meta Self-Learning for Multi-Source Domain Adaptation: A Benchmark},
  author={Qiu, Shuhao and Zhu, Chuang and Zhou, Wenli},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={1592--1601},
  year={2021}
}

License

This Dataset is made freely available to academic and non-academic entities for non-commercial purposes such as academic research, teaching, scientific publications, or personal experimentation. Permission is granted to use the data given that you agree to our license terms bellow:

  1. That you include a reference to our paper in any work that makes use of the dataset.
  2. That you may not use the dataset or any derivative work for commercial purposes.

Privacy

Part of the data is constructed based on the processing of existing databases. Part of the data is crawled online or captured by ourselves. Part of the data is newly generated. We prohibit you from using the Datasets in any manner to identify or invade the privacy of any person. If you have any privacy concerns, including to remove your information from the Dataset, please contact us.

Contact

Reference