yongliu20 / UniLSeg

[CVPR 2024] Official implementation of "Universal Segmentation at Arbitrary Granularity with Language Instruction"

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

I get poor result

Kaisor-Yuan opened this issue · comments

Is there anything I need to adjust?
Why is the result I get different form yours so much?
e94387e1e6219e0cb39d00dc5e87165
f28538a7ad65851e8ce0a2100afab36

What's your input instruction? I also test this image with the released ckpt and get expected segmentation results. Just as the following shows (the instruction is "all light" and "all picture", respectively). Thus I think your input instruction or something is wrong. Besides, the visualization results in the paper are generated by UniLSeg-100 (pretrained on all SA-1B) and the model is finetuned on ADE20K as well. The released UniLSeg-20 without finetuned may perform a little worse.
all_light
all_picture

My instruction is all pillow

Here is the result I tested.
all_pillow

oh, sorry!
I find that when I download the UniLSeg-20, there is something wrong.
Now I get the satisfactory result.
Thanks for you great work!

Happy to hear that!