ZhengPeng7/BiRefNet

No pth checkpoint for HRSOD

Closed this issue · 2 comments

Hello, thank you for your work, can not find pretrained weights for Sailent Object Detection, the gdrive contains only checkpoints for DIS and COD. Could you please provide well-trained weights on HRSOD too.

Sure. But as I mentioned in README, I've been recently re-training my model on only a single A100-40G card.
And training in HRSOD actually costs the most computation, since there are many settings and more images. It's still in training. But I just got the final results in the setting of using DUTS (D) as the only training set. Well-trained weights and predicted maps are being updated in google-drive, you can download them in 10 mins. And results are even better than the previous ones, which I will update in the next version of our paper:

+----------+-----------+----------+-------+--------+------+-------+--------+-----------+-------+-------+------+
& Dataset  &   Method  & Smeasure & maxFm & meanEm & MAE  & maxEm & meanFm & wFmeasure & adpEm & adpFm & HCE  &
+----------+-----------+----------+-------+--------+------+-------+--------+-----------+-------+-------+------+
& DAVIS-S & tm--ep110 &   .965   &  .964 &  .983  & .008 &  .992 &  .937  &    .937   &  .975 &  .905 & .000 &
& DUT-OMRON & tm--ep110 &   .870   &  .820 &  .884  & .043 &  .892 &  .806  &    .797   &  .884 &  .795 & .000 &
& TE-DUTS & tm--ep110 &   .938   &  .932 &  .956  & .020 &  .965 &  .912  &    .910   &  .937 &  .871 & .000 &
& TE-HRSOD & tm--ep110 &   .956   &  .956 &  .973  & .014 &  .982 &  .936  &    .929   &  .963 &  .911 & .000 &
& TE-UHRSD & tm--ep110 &   .932   &  .935 &  .946  & .028 &  .954 &  .921  &    .906   &  .925 &  .890 & .000 &
+----------+-----------+----------+-------+--------+------+-------+--------+-----------+-------+-------+------+

thank you very much