xuebinqin / DIS

This is the repo for our new project Highly Accurate Dichotomous Image Segmentation
Apache License 2.0
2.28k stars 270 forks source link

U2Net accuracy performs better then DIS on almost every image #37

Open roimulia2 opened 2 years ago

roimulia2 commented 2 years ago

Hey guys, loved the paper!

If I understand correctly, DIS should be the successor of U2Net. In the tests I've made seems like DIS delivers higher quality mask yet less accurate result when comparing it to U2Net. I benchmarked both of them and used DIS V2 general use model.

Is this the expected behavior?

xuebinqin commented 2 years ago

Thanks for the interest. we haven't release the v2 model yet, the model for general use is still v1. The training set of u2net and isnet is different because the high resolution training set is now still limited to certain categories, please refer to the paper for the exact categories. Our v2 dataset and model will try to solve the problem.

On Wed., Aug. 31, 2022, 3:24 a.m. roimulia2, @.***> wrote:

Hey guys, loved the paper!

If I understand correctly, DIS should be the successor of U2Net. For tests I've made seems like DIS delivers higher quality mask yet less accurate result when comparing it to U2Net. I benchmarked both of them and used DIS V2 general use model.

Is this the expected behavior?

ā€” Reply to this email directly, view it on GitHub https://github.com/xuebinqin/DIS/issues/37, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADSGORLYSYWLUTDB5FYNZPDV34XENANCNFSM6AAAAAAQBHPHTE . You are receiving this because you are subscribed to this thread.Message ID: @.***>

roimulia2 commented 2 years ago

Hey @xuebinqin, thanks for the fast response. What is the expected ETA of V2?

roimulia2 commented 2 years ago

Hey @xuebinqin , been almost 2 months. Any chance you are going to release the V2? I'm excited to use it in my university seminar paper!

DanielZanchi commented 1 year ago

Following

iborko commented 1 year ago

We recently tested a number of open-source solutions for our use-case (fashion-related images), and DIS general-purpose model was better than the U2Net model but still worse than some commercial solutions.

We liked the DIS model and are waiting for the v2 version as well. It seems that the biggest issue is the dataset and DIS will be probably comparable to commercial solutions when it gets a larger dataset (maybe some other universities can help?).

You can read our comparison here: https://www.velebit.ai/blog/background-removal/

DanielZanchi commented 1 year ago

@iborko Very interesting article. I also saw PhotoRoom but the pricing is very high. Do you think it would be possible to do a training and improve DIS?

iborko commented 1 year ago

@DanielZanchi Yes, definitely. But DIS authors will know more about that.

lylinelim commented 1 year ago

I also saw PhotoRoom but the pricing is very high

Hello @DanielZanchi, Iā€™m Lyline from PhotoRoom šŸ‘‹ Would you be open to discuss about what would be a better price point from your point of view? Our API product is quite new so happy to get feedback šŸ™‚

mikebilly commented 10 months ago

@lylinelim Hi, can you share which arch PhotoRoom uses, does it use U2Net?. If so what model does it use?