Batch_normalization in encoder part

[elzeru]

Hi, compared to Joker's code, you added Batch_normalization in encoder part. Does it improves the performances ? I tried quickly and it seems to me that it does not produce better results. What do you think of it ?

[protossw512]

The quick answer is I don't know.

Well, the general goal of batch normalization is not to improve performance, but better convergence rate, and maybe some extend of regularization.
Joker's code has some issue that cannot reproduce the results in paper, I modified some part of it and added something. When I ran experiment I did not compare with/without BN so I cannot tell which is better.
But that is a good point, I might try that as well.

[elzeru]

Ok, I just made a quick try and normalizations in the encoder is does not seem to improve convergence, but seriously increases the memory use. I add to reduce my batch size (I can usually use 5-6 images).
Concerning the performances, I reached very promising results on the alphamatting test cases (troll, pineapple,...). The key is to feed the network with big images during the prediction (320x320 is okay for training). For instance, good results can be obtained on Troll with 2560x2560 tiles and about 200 pixel of overlapping.

[protossw512]

Cool! glad to hear that, I actually use this model to so something else so I will probably give it a try. Thanks for sharing your results.

[elzeru]

You're welcome !

[aigc-boys]

@elzeru ,Can you send me the Adobe datasets, I have send a email for author of Deep Image Matting, but I didn't get a reply.Thanks. My email is 1575262785@qq.com.

[elzeru]

Hi, Adobe is giving its set to Academics only so I didn't get it.
However, you can find a lot of png with alpha channel on the web. For the background, there is massive datasets used for segmentation training.