Can someone share an tips to train a better model for text detection and segmentation ? I'm also willing to throw a $100 bucks at the problem

[tonyfuyung]

Can someone share an tips to train a better model for text detection and segmentation ? I'm also willing to throw a $100 bucks at the problem

[ogkalu2]

I'm currently training a model that will hopefully be a lot better or a lot more robust.

The biggest bottleneck is the dataset. The last models were trained on mostly manga and a sizable number of webtoons. Maybe ~200 from the west. Now it's about 2.5k.

Training itself is not that expensive. You can rent instances on vast.ai for a few hours that won't cost more than a few dollars.

[ogkalu2]

Hi. Yes. I'm trying to train multiple architectures and seeing what fares best. A donation would be appreciated.
https://ko-fi.com/comictranslate

[tonyfuyung]

damn the 100usd mastercard gift birthday card I got doesn't work on kofi. I'll need to see how to work around that or swap it for cash with someone.

I'll donate soon!

[tonyfuyung]

@ogkalu2 okay i sponsored you on github for 100usd. (kofi didn't work for my giftcards)
good luck man!

if one day you're willing to shared the code for training or image datasets , I'd love to help out and learn too!

[gullyswander]

I'm currently training a model that will hopefully be a lot better or a lot more robust.

The biggest bottleneck is the dataset. The last models were trained on mostly manga and a sizable number of webtoons. Maybe ~200 from the west. Now it's about 2.5k.

Training itself is not that expensive. You can rent instances on vast.ai for a few hours that won't cost more than a few dollars.

Is there a timeline for that ? do you need more gpu's ? or data ?

What would speed up the training stuff ? New to LLM/VISION AI, not to coding.

[ogkalu2]

@tonyfuyung
Thanks. I don't mind sharing the code (and the parts of the new dataset i can share) when i'm all done. The old one is on roboflow here

• https://universe.roboflow.com/speechbubbledetection-y9yz3/bubble-detection-gbjon
• https://universe.roboflow.com/textbubbledetection/comic-text-detection

@gullyswander
There should be an update within a week or so. I rent on vast so nothing would really speed things up at this point. I don't think i need more data at the moment. The train dataset is about 11k images. One model will detect bubbles and text unlike previously.

[ogkalu2]

@tonyfuyung @gullyswander
The update for the new detection model is out