![]() ![]()
If you meant to do this, you must specify 'dtype=object' when creating the ndarray ![]() Python 3.6.9 VisibleDeprecationWarning: Creating an ndarray from ragged nested sequences (which is a list-or-tuple of lists-or-tuples-or ndarrays with different lengths or shapes) is deprecated. #E Package Apache2 Has No Installation Candidate pdfSee the pdf file sent over slack to reproduce. The dilation factor is set to 1.4, which often corresponds to an overlap of 2 characters, depending on the font.Īfter checking, it does slow down the inference speed on CPU of less than 7% on documents with many boxes to split, and of less than 1% on documents with few boxes to split.Įnhancement module: models topic: text recognition Thus it is robust to long repetitions, such as MRZ codes. It takes a dilation factor as parameter, corresponding to the geometrical overlap between samples of one word, which is used only to determine the number of occurrences of a character when the word is splitted on a repetition of more than 2 characters. Levestein distance between the prefixes/suffixes). #E Package Apache2 Has No Installation Candidate fullWe propose a merge_sequences function in the recognition.utils module, which is smart enough to reconstruct a full word from its samples (computing the min. The splitted crops are dilated so they overlap, thus it is easy to merge back full words with a high precision. The splitting is activated by default when aspect ratio is > 8, which corresponds to 50% of the image padded (aspect ratio of 4 in the model). This should prevent the predictor to be fed with very long (and thus highly vertically padded) sequences, which was leading to very poor result. ![]() This PR adds a crop splitting feature in the recognition predictor to split too long boxes in small ones before feeding the recognition model. GPU models and configuration: GPU 0: GeForce RTX 2060ĬuDNN version: Probably one of the following: I expected the weights to be loaded properly Environment DocTR version: 0.3.0a0 #E Package Apache2 Has No Installation Candidate zip fileZipfile.BadZipFile: File is not a zip file Raise BadZipFile("File is not a zip file") Yields: Traceback (most recent call last):įile "/home/laptopmindee/doctr/text.py", line 8, in įile "/home/laptopmindee/doctr/doctr/models/utils/tensorflow.py", line 50, in load_pretrained_paramsįile "/usr/lib/python3.8/zipfile.py", line 1269, in _init_įile "/usr/lib/python3.8/zipfile.py", line 1336, in _RealGetContents Master = recognition.MASTER(vocab=VOCABS, input_shape=(32, 128, 3)) To Reproduce from doctr.models import recognitionįrom import load_pretrained_params I can't load MASTER weights in TF with the load_pretrained_params function. The End to end ocr named doctr developed by you is fantastic.It is very easy to use and have very good results.Currently i am working on ocr related project.I had implemented doctr on sample images and have received good results.However I had few question which i list below and would be grateful for receiving explanatioins on them.ġ)Which dbresnet50 model are you using?pretrained on synthtext dataset or tested on real world dataset as mentioned in the paper?ģ)Is their anyway we can get output after detection without postprocessing?Ĥ)how can we improve accuracy of detection?ĥ)when would your private dataset be available?Ħ)How much training data we need to get good results on our dataset?(dataset type would be forms,invoices,receipts etc)ħ)Also you have mentioned that to train the model Each JSON file must contains 3 lists of boxes.Why 3 boxes are needed for single image? question ![]()
0 Comments
Leave a Reply. |
Details
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |