How to improve the mAP on my own dataset

[niluanwudidadi]

Hi, @zhreshold I know you have improved the mAP of pascalvoc dataset by multiple modifications , now I have met a problem , when I use ssd, the accuracy on my own dataset is no better than faster-rcnn, I want to know how to improve the performance . In your code ,I have seen many data augument policy, I want to konw is there necessityto open it, and in your code, there are some parameters. I don't konw the label-width's function. Can you give me a explanation? Thank you very much!

[niluanwudidadi]

And the pretrained model is just a vgg_reduced? Is there are better pretrained model?

[niluanwudidadi]

I'm doing character detection, the number character mAP is very low ,just 0.90 around, but Chinese characters and letter characters is very high. Is there any way to improve it?

[zhreshold]

balancing class examples is important, I guess your number characters are very few compared to others.

[niluanwudidadi]

Thank you, I will have a try. @zhreshold