CRF post processing

[xuebinqin]

CRF post processing

[xuebinqin]

I don't suggest to use CRF as the post-processing, because it often introduces many salt-pepper noises along the boundaries. If you want sharp and clear boundaries, you can try our previous BASNet (https://github.com/NathanUA/BASNet) or use the cascaded PSP (https://github.com/hkchengrex/CascadePSP) as the post-process of the U^2-Net results.

[PytaichukBohdan]

@tiwari93 please keep updated if you manage to fine-tune the model.
I've tried:

• fine-tuning with the custom dataset (300 images), but the results became poor after first iterations and haven't improved
• fine-tuning on the original dataset + our custom. The results still became worse

Also, I tried different configurations of layer freezing, with different parameters of Adam optimizer.
But couldn't produce good results

[tiwari93]

Thank u for your response and suggestion.

I will try these and share you the results.

[tiwari93]

Hi,

I just wanted to share few update.
I tried cascade psp for post processing of u2net result and results are really good but it requires GPU and also it take too much time. So here we are losing the benefit of fast processing of u2net.
I just have one thought, Is it possible to fine tune u2net pre trained model on our custom dataset.
Please suggest and thank you again for ur help

[vincentzhang]

It might be easier to train from scratch on your dataset than fine-tuning