Training my own model

[mazatov]

I'm trying to train my own model that recognizes the 1-2 digit numbers. I have my own dataset with label and bounding boxes around each number. I tried training using the default det_mv3_db.yml config file but the text detection results were surprisingly bad.

So I assume I need to change a few things in the config file, especially in the Train and Test groups. The main difference from the ICDAR is that my images are pretty small. (50-100 pixels). So I'm guessing some parameters need to change. Are there any guidelines on how to set up config file for you own dataset?

[LDOUBLEV]

It would be great if you could post a sample from your dataset.

If the number area is obvious in the image, then there is no need to train a text detection model to detect text. It is recommended to train a text recognition model directly.
example：
https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.6/doc/imgs_words_en/word_401.png

[gavinfaimdata]

I just tried resizing all the images in my preprocessing step to 640^2. Hopefully it solves this issue. If not I'll try making the adjustments to the config file

[gavinfaimdata]

To get good results I had to resize the image in main() in predict_system.py

[MrAsifKhan]

@gavinfaimdata did you get good results when you trained the detection model with 640x640??
I am trying to train ResNet50 with SAST with image size of 128*128 and i am getting really bad results.

[gavinfaimdata]

I ended up using 320*320 in the end and the results are decent... not as good as I want.

[camcomharsh]

increase your resize to [960, 960] and add more augmentation as you are already using mobile net architecture and if possible you should try exploring resnet based architectures