Do color feature affect accuracy? about calamari HOT 5 CLOSED

It is not surprising that a model trained on white characters and black background cannot be used to predict the inverse. The model can only predict what it knowns: Usually this should be black foreground and white background in the case of Calamari.
Calamari does not support color features, since it converts all data to grayscale. However, you need to train on gray-scale images in order to receive meaningful results.

The actual prediction time for predicting a single line are about 0.3 Seconds per line on a single CPU. I guess most of the time is needed to load the model, however this must only be done once. Thus, to reach these times, you must manually create and store a calamari_ocr.ocr.Predictor-Object and initialize it. Then use this model to predict a line:
E. g. (but untested)

from calamari_ocr.ocr import Predictor, create_dataset, DataSetType, DataSetMode

# Load (only once)
predictor = Predictor('PATH_TO_THE_MODEL_WITHOUT_EXT')

# Predict
data = create_dataset(
  DataSetType.RAW,  # if you already have np.arrays or DataSetType.File to load from the filesystem
  DataSetMode.PREDICT,
  images=[raw_image_or_path_to_image],
)
prediction, _ = list(predictor.predict_dataset(data))[0]

from calamari.

thank you for response.

from calamari.

well，i've already create a data like this :
data=FileDataSet(DataSetMode.PREDICT,input_image_files) or
data=RawDataSet(DataSetMode.PREDICT,images)
the first one works well,input_image_files=path to the images(str)
my question is about the second one, the description of the parameter images is the list of images,Is that mean the np.arrays of the image or the list of np.arrays?

from calamari.

images is a list of np.array's for a RawDataSet

from calamari.

that's great,thank you!!!

from calamari.