search

poppinace / indexnet_matting

(ICCV'19) Indices Matter: Learning to Index for Deep Image Matting
Other
389 stars 90 forks source link

about the evaluation code #3

Closed wrrJasmine closed 5 years ago

wrrJasmine commented 5 years ago

@poppinace hi Thank you for your great work You mentioned you have implemented a python version of the evaluation_code. But I can't find it in the folder. Would you mind to tell me where is it?

wrrJasmine commented 5 years ago

@poppinace I have one more question. In the line 787 of hlmobilenetv2.py, it is the first layer of mobilenetv2 and you set the stride 1 followed with a maxpooling layer instead of stride=2, so that there is one more unpooling layer in the decoder part. Is it matters to the final results?

poppinace commented 5 years ago

Hi @wrrJasmine ,

The python evaluation code in in the scripts/utils.py

As for the second question, the maxpooling and unpooling layers in mobilenetv2 are introduced to make a direct comparison to the SegNet-based Deep Matting. It matters to the final results. In paper we have reported the numeral results when unpooling is disabled. Please have a look.

wrrJasmine commented 5 years ago

@poppinace I got it. Thank you so much!

wrrJasmine commented 5 years ago

@poppinace hi！ I found that one of the trimap in my test forder is much more different from the other 19 trimaps. the name is "girl-1467820_1280_12.png". It is all gray. I am not sure if it is my fault to mix it. would you mind to check it in your test folder?

poppinace commented 5 years ago

I also noticed this issue. The testing trimaps are not generated by code but are provided by the original authors. Maybe you can drop them an e-mail enquiring this.