Confusion about whether paddle transforms in config autoscales bounding boxes

[fishjojo1]

Hi All,
I am attempting to finetune the PPOCRv5_server_det model
My input images have size 21184x3455
I have tried to remove some augmentations present in the default config

transforms:
    - DecodeImage:
        img_mode: BGR 
        channel_first: false
    - DetLabelEncode: null
    - CopyPaste: null
    - IaaAugment: null
        # augmenter_args:
        # - type: Fliplr
        #   args:                               
        #     p: 0.5
        # - type: Affine
        #   args:
        #     rotate:
        #     - -10
        #     - 10
        # - type: Resize
        #   args:
        #     size:
        #     - 0.5
        #     - 3
    # - EastRandomCropData:
    #     size:
    #     - 960
    #     - 960
    #     max_tries: 50
    #     keep_ratio: true]
    # - RecResizeImg:
    #     image_shape: [3, 48, 320]
    - Resize:
        target_size: [960, 960]
        keep_ratio: true
        padding: true

However, it seems as if the bounding boxes are not being scaled with the images as I am getting errors like

File "/PaddleOCR/ppocr/data/simple_dataset.py", line 137, in __getitem__
    outs = transform(data, self.ops)
  File "/PaddleOCR/ppocr/data/imaug/__init__.py", line 73, in transform
    data = op(data)
  File "/PaddleOCR/ppocr/data/imaug/make_border_map.py", line 59, in __call__
    self.draw_border_map(text_polys[i], canvas, mask=mask)
  File "/PaddleOCR/ppocr/data/imaug/make_border_map.py", line 114, in draw_border_map
    canvas[ymin_valid : ymax_valid + 1, xmin_valid : xmax_valid + 1] = np.fmax(
ValueError: could not broadcast input array from shape (37,0) into shape (37,1)

With EastRandomCropData, this is fixed but my loss flatlines at 1 even after many epochs.
May I verify that I am setting up training right?
This is what I am doing

1. Have a list of images data/image1.jpg....image100.jpg
2. each image has a few lines of text
3. extract the text and use gen_label.py to produce the following
   <image_file> [{"transcription": "xxxxx", "points": [[125, 79], [546, 79], [546, 93], [125, 93]]}, {"t
4. run train.py

[liuhongen1234567]

Hello, it seems that the issue is caused by your image being too large while the text box is too small. For example: [[125, 79], [546, 79], [546, 93], [125, 93]], its height is 14, and the original image height is 3455. If you scale it down to 960, the scaled bounding box height would be approximately 5 pixels, which makes it difficult for the model to converge. You can try using a larger scaling size, such as Resize target_size: [2000, 2000].