Self-supervised text erasing

Dataset

The example datasets are provided in './examples' . And the entire datasets can be found as follows:

PosterErase

You can download our poster text datatset at PosterErase.
SceneErase

You can download the scene text datatset at SCUT-EnsText. And the extended unlabeled scene text dataset can be found at ICDAR2019-ArT.

Model

You can download the checkpoints for the posterErase dataset as follows:

Self-supervised Text Trasing :

download from oss, and put it under './checkpoints/erasenet/ste/best_net_G.pth'
Finetuning on STE :

download from oss, and put it under './checkpoints/erasenet/ft/best_net_G.pth'

Runing

test

python test.py --dataset_mode items --dataroot ./examples/poster --model erasenet --name ft --which_epoch best  # inferece with the ste model on poster 

python test.py --dataset_mode items --dataroot ./examples/poster --model erasenet --name ste --which_epoch best # inferece with the finetuned model model on poster

train offline

python train.py --batchSize 2 --dataset_mode items --dataroot ./examples/poster --model erasenet --valid 1  --gen_space random8 --name item-rand-s8-3 --PasteImage --sigmoid     # train with random synthesis policy on poster 

python train.py --batchSize 2 --dataset_mode ens --dataroot ./examples/scene --model erasenet --valid 1  --gen_space random8 --name ens-rand-s8-3 --PasteImage --mask_sigmoid --sigmoid # train with random synthesis policy on erase

train online with policy update

python train_adg.py --batchSize 2 --dataset_mode items_adg --dataroot ./examples/poster  --model erasenet  --name item-sp8-r24 --gen_space random8 --lambda1 1 --lambda3 0 --lambda2 5 --reward_norm exp --reward_type 24 --valid 1 --ctl_M 3 --ctl_freq 10 --ctl_policy_num 2 --ctl_train_freq 10 --ctl_batchSize 2 --ctl_update_num 6  --PasteImage --sigmoid  # train with self-supervised synthesis policy on poster 

python train_adg.py --batchSize 2 --dataset_mode ens_adg --dataroot ./examples/scene  --model erasenet  --name ens-sp8-r24 --gen_space random8 --lambda1 1 --lambda3 0 --lambda2 20 --reward_norm exp --reward_type 24 --valid 1 --ctl_M 3 --ctl_freq 10 --ctl_policy_num 2 --ctl_train_freq 10 --ctl_batchSize 2 --ctl_update_num 6  --PasteImage --mask_sigmoid --sigmoid # train with self-supervised synthesis policy on erase

train online with policy outputed by feature

python train_adg_data.py --batchSize 2 --dataset_mode items_adg_fea --dataroot ./examples/poster  --model erasenet --gen_space random8 --name item-sp8-data-r24 --clr 0.00005  --reward_norm exp --ctl_layer 2 --reward_type 24 --lambda1 1 --lambda2 10  --valid 1  --ctl_freq 10  --ctl_batchSize 2 --ctl_update_num 10 --PasteImage --sigmoid # train with self-supervised synthesis policy on poster 

python train_adg_data.py --batchSize 2 --dataset_mode ens_adg_fea --dataroot ./examples/scene  --model erasenet --gen_space random8 --name ens-sp8-data-r24 --clr 0.00005  --reward_norm exp --ctl_layer 2 --reward_type 24 --lambda1 1 --lambda2 10  --valid 1  --ctl_freq 10  --ctl_batchSize 2 --ctl_update_num 10 --PasteImage --mask_sigmoid --sigmoid # train with self-supervised synthesis policy on erase

预处理json文件内容

./examples/poster/train.txt
{
    "mask": "x1,x2,y1,y2;x1,x2,y1,y2;... ..." ;
    "place": {
        "obj": "text,size,direction;text,size,direction;... ..." ;
        <!-- "raw": "x1,x2,y1,y2;x1,x2,y1,y2;... ..." ; -->
        "text": [
            [
                [x,y,["c1,c2,c3",... ...]], 
                [x,y,["c1,c2,c3",... ...]], 
                ... ... 
            ],
            [
                [x,y,["c1,c2,c3",... ...]], 
                [x,y,["c1,c2,c3",... ...]], 
                ... ... 
            ], ... ...
        ]
    }
}

Acknowledge