search

Yaoyi-Li / GCA-Matting

Official repository for Natural Image Matting via Guided Contextual Attention
MIT License
392 stars 62 forks source link

About the performance #4

Open hengshan123 opened 4 years ago

hengshan123 commented 4 years ago

Thank you for the excellent work, i use your pretrained model gca-dist and test in composition-1k dataset, but i get the sad, mse like bellow, it's a little different from the paper [03-12 15:15:46] INFO: TEST NUM: 1000 [03-12 15:15:46] INFO: MSE: 0.010313104958540355 [03-12 15:15:46] INFO: SAD: 37.83985428515623

Yaoyi-Li commented 4 years ago

I checked every steps again and it works fine for me. [03-13 02:22:57] INFO: TEST NUM: 1000 [03-13 02:22:57] INFO: MSE: 0.009076512024573515 [03-13 02:22:57] INFO: SAD: 35.278531444580096

Did you generate merged images by the code Composition_code.py provided in Adobe dataset and follow their image order test_bg_names.txt/test_fg_names.txt?

xiezhang666 commented 4 years ago

Thanks for your excelent job. I use your pretrained model gca-dist and test in composition-1k dataset as hengshan123's talk. here is my result:

['model-600238_1920_13.png'] 8.949408203125 [04-08 18:23:52] INFO: TEST NUM: 1000 [04-08 18:23:52] INFO: MSE: 0.010313099013678589 [04-08 18:23:52] INFO: SAD: 37.83984849536125

I merged Adobe dataset following huochaitiantang's sciprt. (https://github.com/huochaitiantang/pytorch-deep-image-matting/blob/master/tools/composite.py)

Yaoyi-Li commented 4 years ago

Thanks for your excelent job. I use your pretrained model gca-dist and test in composition-1k dataset as hengshan123's talk. here is my result:

['model-600238_1920_13.png'] 8.949408203125 [04-08 18:23:52] INFO: TEST NUM: 1000 [04-08 18:23:52] INFO: MSE: 0.010313099013678589 [04-08 18:23:52] INFO: SAD: 37.83984849536125

I merged Adobe dataset following huochaitiantang's sciprt. (https://github.com/huochaitiantang/pytorch-deep-image-matting/blob/master/tools/composite.py)

I tried this script and I can reproduce your result. Maybe hengshan123 also used this one. But it is different from how Adobe merge their Composition-1k testing set. So don't use this code to merge your data for testing. I have uploaded the composition_code.py and it uses the composition function provided by Adobe. You can try it.