README.md

BlendMask instance detection

coco/
  thing_train2017/
    # thing class label maps for auxiliary semantic loss
lvis/
  thing_train/
    # semantic labels for LVIS

Run python prepare_thing_sem_from_instance.py, to extract semantic labels from instance annotations.

Run python prepare_thing_sem_from_lvis.py, to extract semantic labels from LVIS annotations.

Text Recognition

English pretrained data:

• Totaltext training, testing images, and annotations [link] [paper] [code].
• CTW1500 training, testing images, and annotations [link] [paper] [code].
• MLT [dataset] [paper].
• Syntext-150k:
  • Part1: 94,723 [dataset]
  • Part2: 54,327 [dataset]
  • If you have trouble downloading Syntext-150k, you can try BaiduNetDisk [here]

Chinese pretrained data:

• ReCTs [images&label](1.7G) [Origin_of_dataset]
• ReCTs test set [images&empty_label](0.5G) [Origin_of_dataset]
• LSVT [images&label](8.2G) [Origin_of_dataset]
• ArT [images&label](1.5G) [Origin_of_dataset]
• SynChinese130k [images&label](25G) [Origin_of_dataset]

text/
  totaltext/
    annotations/
    train_images/
    test_images/
  mlt2017/
    annotations/train.json
    images/
    ...
  syntext1/
  syntext2/
  ...
  evaluation/
    gt_ctw1500.zip
    gt_totaltext.zip

To evaluate on Total Text and CTW1500, first download the zipped annotations with

mkdir evaluation
cd evaluation
wget -O gt_ctw1500.zip https://cloudstor.aarnet.edu.au/plus/s/xU3yeM3GnidiSTr/download
wget -O gt_totaltext.zip https://cloudstor.aarnet.edu.au/plus/s/SFHvin8BLUM4cNd/download

Person In Context instance detection

pic/
  thing_train/
    # thing class label maps for auxiliary semantic loss
  annotations/
    train_person.json
    val_person.json
  image/
    train/
    ...
  

First link the PIC_2.0 dataset to this folder with ln -s \path\to\PIC_2.0 pic. Then use the python gen_coco_person.py to generate train and validation annotation jsons.

Run python prepare_thing_sem_from_instance.py --dataset-name pic, to extract semantic labels from instance annotations.