search

PeterL1n / BackgroundMattingV2

Real-Time High-Resolution Background Matting
MIT License
6.85k stars 952 forks source link

Has the comparison algorithm been retrained? #101

Closed rgbd-zml closed 3 years ago

rgbd-zml commented 3 years ago

You compared DIM and FBA algorithms. Have you retrained them on the VideoMatte240K, PhotoMatte13K, and Distinctions-646 data sets. Or directly use the pre-trained model provided by the original author to test.

PeterL1n commented 3 years ago

No. Only BackgroundMattingV1 were retrained for comparison

rgbd-zml commented 3 years ago

No. Only BackgroundMattingV1 were retrained for comparison Thank you for your reply. If the training sets are different, might it be unfair? What do you think

PeterL1n commented 3 years ago

No didn't have time to train every methods. Additionally, FBA uses trimap, we already acknowledged with gt trimap, their method does better. It was a different category of method and not the focus of our study.

rgbd-zml commented 3 years ago

I have an idea, if DIM is jointly trained on three data sets. Can its indicators be better? DIM may also exceed your indicators. In other words, adding a background image can't exceed the existing trimap-based algorithm? Anyway, BGV2 is still a great job.