Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[Question] Configuring the dataset for custom training #409

Open
kmaru0702 opened this issue Oct 25, 2023 · 2 comments
Open

[Question] Configuring the dataset for custom training #409

kmaru0702 opened this issue Oct 25, 2023 · 2 comments

Comments

@kmaru0702
Copy link

kmaru0702 commented Oct 25, 2023

I am in the process of setting up a custom dataset for custom training.

Q1. Is it okay to include images that contain words not used in the dataset?
For example, if the words I want to train on are 1,2,3, can I include an image and text with "1ABC2E3" in the dataset?
( like : "1ABC2E3.jpg","1ABC2E3" )

Q2. The default values for imgH and imgW during training are 32x100.
Is it okay to use these values even if the aspect ratio of the training image is 1:1?

For example, when the training image is roughly 64x70, which is closer to a square.

Q3. If I want to keep the numbers as they are and only personally add characters,
is it acceptable to only add those specific character images?

Q4. use preprocessing img for train is better?

@kmaru0702 kmaru0702 closed this as not planned Won't fix, can't repro, duplicate, stale Oct 25, 2023
@Seoung-wook
Copy link

Let me explain what I know. I hope this helps.

Q1. you should edit your character.(Please refer to --> parser.add_argument('--character', ....))
Q3. The input image must be a cropped image of the text area, and the label must be text corresponding to the area.

@kmaru0702
Copy link
Author

@Seoung-wook Thanks a lot

@kmaru0702 kmaru0702 reopened this Oct 27, 2023
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants