search

ZhengPeng7 / BiRefNet

[CAAI AIR'24] Bilateral Reference for High-Resolution Dichotomous Image Segmentation
https://www.birefnet.top
MIT License
1.26k stars 100 forks source link

Model weights when trained on HQ-Seg44k dataset #51

Closed chrismaltais closed 1 month ago

chrismaltais commented 3 months ago

Hey there!

Segment Anything in High Quality (SAM-HQ) model released a dataset called HQ-Seg44k, which can be found here. It includes the DIS5K dataset, but also other data that might be worthwhile to train on.

We seem to be on a roll of releasing different model weights for different practical applications, I'd be curious to see how including this new dataset in the general use model with the swin_v1_large backbone could affect model quality.

Kind of like what has been done by adding a dataset here

ZhengPeng7 commented 3 months ago

Hey, Chris, thanks for your suggestion! I've downloaded them to have a check, and they seem to be good to put into training 👍. But currently, the GPU support from Fal might be going to be closed (they have offered me so many GPU resources for months, and I really appreciate them). Therefore, I might not have that many computation resources to conduct this kind of massive training again. But I've kept them and your suggestion, which may work in the future.