deepinsight/insightocr

MXNet OCR implementation. Including text recognition and detection.

insightocr's Issues

lstm problem

There's  a line in     lstm.py  .    

last_states.append(LSTMState(c=mx.sym.Variable("l%d_init_c" % i), h=mx.sym.Variable("l%d_init_h" % i)))

I found Variable c & h are not initialized or setted. So why the symbol is right?

Unable to train insightocr on VGG_Text data

Hi,
I am not able to train insightocr using train_crnn.py file.
I have downloaded the Synthetic Word Dataset(10GB) from VGG_Text.
Then i have made following changes-

• In config.py,
  default.dataset = 'vgg'
  dataset.vgg.dataset_path = 'path of vgg dataset'

• In train_crnn.py,
  In line 70- image_set='annotation_train'
  In line 71- image_set='annotation_val'

annotation_train file contains data in following format-

./2425/1/115_Lube_45484.jpg 45484
image_name index

These index are mapped to label in lexicon.txt file. But in the entire script, we have not passed the path of lexicon.txt file. So how the model is going to train.

Kindly provide some solution to train it on VGG_text dataset.

Also what is config.image_path in config.py?

mxnet原生ctc试过吗?

image data prepocess

Hi,

About the code in your crnn/data.py (line 121-122)

_data -= 127.5
_data *= 0.0078125

Could you please explain what you mean by this? Thanks.

这个库考虑对一些具体的场景的 图片 做嘛

比如一些证件的ocr

acc 一直是0

你好，我用你的代码重新训练模型。实验的时候，除了带LSTN的网络，其他不带LSTM的网络，比如simplenet、resnet、mobilenet都试过，训练过程中acc一直都是0，请教一下，这是什么原因，谢谢

Is 4x1pooling better than original?

   I found you have changed the cnn part of crnn  by using 4x1pooling .Is it better?

Pretrained model, detection and inference code

Hi @nttstar,

This project is great.

when all the thing is ready?. thanks