search

lizhengwei1992 / Semantic_Human_Matting

Semantic Human Matting
532 stars 146 forks source link

Question about model performance #19

Closed FantasyJXF closed 5 years ago

FantasyJXF commented 5 years ago

Thanks for your nice work.

I trained your model with aisegment dataset, during the test time, there are always white block on the foreground like this 下载

And I test with your pretrained model, it also have same problem

mask1

Looking forward to your reply....

hassanrazakhalid commented 5 years ago

yes i am also getting the same issue on pretrained model. I have not tried my own model yet. Please see image below https://ibb.co/rQ0xMzd. If some has any solution please share or if some one has achieved better results

FantasyJXF commented 5 years ago

yes i am also getting the same issue on pretrained model. I have not tried my own model yet. Please see image below https://ibb.co/rQ0xMzd. If some has any solution please share or if some one has achieved better results

if you use numpy, just clip the value to [0,255] and use dtype np.uint8 would solve the problem.

The model really need good datasets which was so so hard to get

hassanrazakhalid commented 5 years ago

yes but this only remove white and black, what about greenish background which is merged as foreground

FantasyJXF commented 5 years ago

That's because the pretrained model was not good enough. That's what you could improve.

The picture you provided has green background, did it?

Have you try https://www.remove.bg/?

hassanrazakhalid commented 5 years ago

First link is from remove.bg https://ibb.co/FzKFyx1 This is from your algo and provided model. https://ibb.co/RNyRnxb

The results of remove.bg are very good.

  1. Are they using the same approach that is used in this.
  2. Also as you suggested to improve model. I just need a good data set or i need to change the trainning model as well.
  3. For dataset Is the following dataset good enough https://www.kaggle.com/laurentmih/aisegmentcom-matting-human-datasets