1. • Auto-rotate
2. • Detection cropping
3. • Recognition
4. • Connect all the section
5. • Auxiliary methods (see Recognition Part in TODO)

• 2021/02/25 Results:
  • Rotation Head : 98~100 %
  • YOLO map: 96.8 -> 100
  • Upper part: 92.77 -> 96.39 %
  • Lower part: 92.77 -> 98.80 %
  • Combination: 84.34 -> 90.36~92 %

• data.csv: contains overall annotations
  • col: |ID|file_name|GT_1|GT_2|xmin|ymin|xmax|ymax|ymax_2|mode|
  • GT_1: ground truth of upper region
  • GT_2: ground truth of lower region
  • mode: 0/1/2: training/validation/testing

  (xmin,ymin)  --------------------
                |   upper region   |
                -------------------- (xmax,ymax_2)
                |   lower region   |
                --------------------
                                    (xmax,ymax)
  

• MultiplicativeNoise, RandomBrightness, GaussNoise data augmentation hurt LPRnet accuracy
• Rewrite the decoding of LPRnet helps a lotttttttt!

• rotation model: train_rotation.py
  • data in 'data/20201229/EXT/resize/new_image'
• YOLO : yolov4/train.py
  • data in 'yolov4/VOCdevkit/VOC2007'
• lprnet : train_lprnet.py
  • data in 'data/20201229/EXT/resize/new_image'
  • will use the data.csv to crop the images

• Single image: check rotation part in output_pipeline.py
• Evaluate the performance: check test_rotation.py

• Qualitative Result (bbox on images):Check yolov4/predict.py
• Quantitative Result (Map, F1 and diagrams... etc.):
  1. Modifing (change to your training result) the model path in yolov4/yolo.py
  2. Run yolov4/get_dr_txt.py
  3. Run yolov4/get_gt_txt.py
  4. Run yolov4/get_map.py
  5. Results will be saved in yolov4/results

• Inference and image output is in test_lprnet.py
• Drawing diagrams on validation set is in test_lprnet.py
• Video output of the LPRnet result is in utils.py

• • Use cutmix augmentation (postpone)
• • Shuffle the number sequence
• • Test-time augmentation
• • Use the pattern remove image

• Using SHVM dataset train yolo.

• • Change image resolution
• • Enlarge a bit of annotation in x direction
• • Change to cosine scheduler
• • Use different decoder
• • Move to the upper region
• • Split the data into two sets
  1. One with 9 digits
  2. One with 3 digits + whatever
• • Rotation
  • the rotation model is in 'model.py'
  • the weight is in 'weights/rotation/acc_100.00_loss_0.747.pth'
  • Result:
    • val : 98.21% (56 images)
    • test: 100 % (55 images)
• YOLO
  • • modify the xml files
    • check the code in EDA.ipynb (in "new annotation for yolo")
  • • Put in the YOLO

================================== Decrypted ==================================

• • combine two the regions
• Add the src file to increase the features (-)
• • write the metric
• Refine the trainings (-)
• Use the DFT (V)
  • Apply to the data, check the gain (-)
• kmean the bbox (-)

0. Might need to rotate the images first. In preprocessing.py
1. data_check
2. annotation data add in, also col note added
3. dataframe_creation
4. Below will be save in yourDefineName_EXT_clear_2data_mode_resize.csv ID file_name GT_1 GT_2 xmin ymin xmax ymax ymax_2 mode 0 0 20201229080315_0EXT.bmp 389708 102MV 237 756 1267 1336 1046 0 1 1 20201229080446_0EXT.bmp 389708 207V 257 553 1318 1143 848 1 ...
5. pnet_traindata for pnet data
6. assemble_split assemble image list
7. train Pnet in train_pnet.py
8. onet_traindata
9. assemble_split again
10. train_onet.py
11. MTCNN_evaluation.py can give qualitative result
12. mtcnn_metric.py gives the quantitative resulrt

• 'MTCNN/data_preprocessing/gen_Pnet_train_data.py' (V)
• 'MTCNN/data_preprocessing/gen_Onet_train_data.py' (V)
• 'MTCNN/data_preprocessing/assemble_Pnet_imglist.py' (V)
• 'MTCNN/data_preprocessing/assemble_Onet_imglist.py' (V)

• 'MTCNN/train/Train_Pnet.py' (V)
• 'MTCNN/train/Train_Onet.py (V)

• mtcnn_visual.py (V)
• mtcnn_metric.py (V)