Object-Corner-Detection: Crop an object from four corners

License Plate Detection, ID Detection, Vietnam License Plate Detection
Idea: Corners as keypoints, a Centernet modify version

• Different:
  • Modify from bounding box detection to keypoint detection
  • Add augmentations

~ By Hisiter-HUST ~

Note:

• This repository can apply to:
  • License Plate Detection
  • ID Detection
  • ...

Keywords: Vietnam License Plate Detection, ID Vietnam Detection

1. Set up environment and dependent packages

• Step 1: Clone repository:

git clone https://github.com/hisiter97/Object-Corner-Detection.git

• Step 2: Set up conda environment (require conda is installed, if not, install conda):

conda env create -f environment.yml
conda activate py36_torch1.4

• Step 3: Build DCNv2:

sh build_dcnv2.sh

Note

* Build successfully if the output has line ***Zero offset passed*** or ***Passed gradient***
* If there is any error, try downgrade cudatoolkit  

2. Dataset format

• Prepare custom dataset: DATA.md
• Format data: COCO
• Structure

${Object-Corner-Detection_ROOT}
|-- center
|-- data
`-- |-- plate
    `-- |-- annotations
            |-- train_plate.json
            |-- val_plate.json

        |---images
            |--img_1.jpg
            |--img_2.jpg

3. Training:

• Edit config file at the center/config directory

cd center 
python train.py --config config/plate.yml 

4. Testing:

• Download pretrained model:

  • Vietnam License plate: https://drive.google.com/file/d/1GZKi1F7nKLT0d7uLLGewkm4LknqDS57F/view?usp=sharing
  • ID Vietnam: https://drive.google.com/file/d/1igdw06UTw1SZxGa7JUCB-Vm3YiWl1cpt/view?usp=sharing

• Edit line weight in file config.yml

cd center
python detect.py --config config/plate.yml --image_path ../img_test/plate.jpg 

5. DEMO

TODO

• Generate point annotation from testing
• Detect object from 2,3 points
• Debugger
• Invisible keypoint