Quick Start

1.環境準備
2.套件安裝
3.STD
4.STR
5.STD+STR

1. 環境準備

作業系統
- Ubuntu 18.04.6
環境介紹
- Python 3.8.8
- GPU RTX2080 * 1 + V100 * 1
- CUDA Version 10.2

2. 套件安裝

    # Linux example
    conda create --name STDR 
    source activate STDR

    # clone repo
    git clone https://github.com/D641593/STDSTR.git
    cd STDSTR

    # this installs the right pip and dependencies for the fresh python
    conda install ipython pip

    # python dependencies
    pip install -r requirement.txt

    # install PyTorch with cuda-10.2
    conda install pytorch torchvision cudatoolkit=10.2 -c pytorch

    # make sure your cuda path of $CUDA_HOME is the same version as your cuda in PyTorch
    # you need to delete the build directory before you re-build it.
    echo $CUDA_HOME
    # if $CUDA_HOME is not your cuda path, you have to define cuda path
    # >> export CUDA_HOME='/usr/local/cuda'  
    cd STD/assets/ops/dcn/
    python setup.py build_ext --inplace

3. Scene Text Detection ( STD )

# 此階段指令皆須在STD資料夾下執行
cd STD

STD 資料集
請建立datasets資料夾

mkdir datasets

須在STD/datasets/logoHigh/放置訓練資料集，架構如下

|-logoHigh
   |- train_list.txt
   |- test_list.txt
   |- train_images
       |- img_1.jpg
       |- img_2.jpg
       |- img_3.jpg
       | ...
   |- train_gts
       |- img_1.txt
       |- img_2.txt
       |- img_3.txt
       | ...
   |- test_images
       |- img_1.jpg
       |- img_2.jpg
       |- img_3.jpg
       | ...
   |- test_gts
       |- img_1.txt
       |- img_2.txt
       |- img_3.txt
       | ...

資料集下載如下
datasets_STD

訓練
訓練時需要編寫yml檔案，定義訓練時的架構與訓練資料

    yml檔案名稱                 
experiments/seg_detector/base_logoHigh.yaml                 ->  定義資料集
experiments/seg_detector/logoHigh_resnet152_interd2v2.yaml  ->  定義架構

定義完成後，可使用以下指令開始訓練

# example
CUDA_VISIBLE_DEVICES=0 python train.py experiments/seg_detector/logoHigh_resnet152_interd2v2.yaml

訓練完成後模型儲存在STD/workspace/SegDetectorModel-seg_detector/resnet152/L1BalanceCELoss/model

偵測
已訓練模型參數如下
STD trained models
請建立weight資料夾，將下載模型參數放置於此資料夾

mkdir weight

偵測可單張圖片或一整個資料夾，須給定yaml檔案和模型參數檔案

# 單張圖片
CUDA_VISIBLE_DEVICES=0 python demo.py experiments/seg_detector/logoHigh_resnet152_interd2v2.yaml --image_path datasets/logoHigh/private/img_21000.jpg --resume weight/model_epoch_474_minibatch_810000 --box_thresh 0.5 --thresh 0.5
# 資料夾 
CUDA_VISIBLE_DEVICES=0 python demo.py experiments/seg_detector/logoHigh_resnet152_interd2v2.yaml --image_path datasets/logoHigh/private --resume weight/model_epoch_474_minibatch_810000 --box_thresh 0.5 --thresh 0.5
# 輸出帶框圖片 --visualize
CUDA_VISIBLE_DEVICES=0 python demo.py experiments/seg_detector/logoHigh_resnet152_interd2v2.yaml --image_path datasets/logoHigh/private --resume weight/model_epoch_474_minibatch_810000 --box_thresh 0.5 --thresh 0.5 --visualize

輸出位置為 STD/demo_results

4. Scene Text Recognition ( STR )

# 此階段指令皆須在STR資料夾下執行
cd ..
cd STR

字典
字典用以表示所有字對應的編號，若有一字典如下
```
漫
髮
型
工
作
室
```
則標記為「漫」的圖片會對應到[0]，而標記為「髮型工作室」的圖片則會對應到[1 2 3 4 5]
本次使用的字典為
```
STR/train_data/myDict.txt
```

STR 資料集
須在STR/train_data/放置訓練資料集，架構如下

|-train_data
   |- train_list.txt
   |- train_crop
       |- img_1_1.jpg
       |- img_1_2.jpg
       |- img_2_1.jpg
       | ...

資料集下載如下
datasets_STR

訓練
須定義train.yaml 已訓練模型參數下載連結如下
STR trained models
```
python train.py --yaml_file train.yaml
```
訓練完成模型儲存於train_models/xxxxxx，xxxxxx資料夾於train.yaml內可定義

5. STD + STR

若須完整的偵測＋辨識，請將圖片放置STD/datasets/private資料夾內，並執行辨識指令

cd STD
# 資料夾 
CUDA_VISIBLE_DEVICES=0 python demo.py experiments/seg_detector/logoHigh_resnet152_interd2v2.yaml --image_path datasets/logoHigh/private --resume weight/model_epoch_474_minibatch_810000 --box_thresh 0.5 --thresh 0.5

執行 STDtoSTR.py，將STD/demo_results內的結果整合成empty.csv，並根據辨識結果切割STD/datasets/logoHigh/private的圖片至STR/train_data/private_high_crop

cd ..
python STDtoSTR.py --res_img_dir STD/demo_results/ --img_dir STD/datasets/logoHigh/private --output_dir STR/train_data/private_high_crop

須定義predict.yaml，可定義

1. 預測圖片資料夾
2. 圖片csv檔
3. 輸出csv檔
4. 模型參數檔
5. 字典

執行predict.py進行預測

cd STR
python predict.py --yaml_file predict.yaml

得到ans.csv，為最終預測結果