How can I improve performance?

[pervin0527]

hello. I currently work in a medical institution. I referenced your code and was able to achieve 84-85% performance. But I don't know how to improve the performance further here. I tried using a different backbone in mmsegmentation, but this too had similar performance. How did you improve performance??

[tungbt-k62]

I suggest some tricks that could help you improve your performance:

• change decoder segformer dim to 768, in paper author said is could improve segmentation performance
• training with large image size
• mosaic augmentation
• test time aug & ensemble model

[pervin0527]

I suggest some tricks that could help you improve your performance:

• change decoder segformer dim to 768, in paper author said is could improve segmentation performance
• training with large image size
• mosaic augmentation
• test time aug & ensemble model

Thanks for your answer. Is it okay if I try the methods suggested and add more if I have any questions?

[tungbt-k62]

I no longer research on segmentation but if you have any questions, just ask, I will answer according to my knowledge

[pervin0527]

I no longer research on segmentation but if you have any questions, just ask, I will answer according to my knowledge

Hello, referring to the advice you gave last time, I increased the image size from 384 to 512, applied mosaic augmentation, and increased the decoder channel to 768, which slightly improved performance. What can I add here?? I was able to achieve up to 87% performance when I additionally applied copy-paste augmentation.