Beom comments

Results 32 comments of


                                            Beom

Questions about Deeplab retrain part

Thank you for your interest in our work. For your first question, DeepLab Large-FOV is equal to DeepLab-V1, and DeepLab-ASPP is equal to DeepLab-V2. The DeepLab-V2 code produces DeepLab-ASPP (ResNet-101)...

Model structure error?

[please check CRNN paper](https://arxiv.org/pdf/1507.05717.pdf)

ImageNet pre-trained weights link

[click here](https://github.com/qjadud1994/Text_Detector/blob/master/Pytorch/weights/readme.md)

very bad result in the real-plate images

Due to overfitting, it is no wonder that it does not work properly in real images. so, I recommend pre-training with synthetic images and fine-tune with real data.

Can you write a document of how to start the project?

Okay. I will update soon.

I'm wondering label 'Z' meaning

Z means padding. I used padding(Z) to equal the total length of the letters.

IOU calculate

I could convert to polygons, but due to the characteristic of the training data, I think IoU by the rectangle is enough.

K.set_learning_phase(0)

In fact, I did not really understand K.set_learning_phase (0). Since there was too much performance difference according to 0 or 1, I fixed it to 0 and did training and...

Bi-LTSM's implementation

I do not really understand your comment. Is it reverse because I used go_backwoards = True in lstm_1b? If it is wrong, please let me know how to fix it.

Bi-LTSM's implementation

Now I understand. I've been using it wrong so far. I modified your `model.py` code with your help. Thank you.