search

dvlab-research / PanopticFCN

Fully Convolutional Networks for Panoptic Segmentation (CVPR2021 Oral)
Apache License 2.0
393 stars 53 forks source link

pretrained weigth name mismatch issue #14

Closed emjay73 closed 3 years ago

emjay73 commented 3 years ago

Hi, first of all, thank you for the inspirational work!

I tried to download the pre-trained weight from the link you provided,

especially the one with PanopticFCN-400 / R50 / 3x / faster threshold.

Download was successful, but the name of the weight is "panoptic_fcn_r50_512_3x.pth", which does not match the concept of it.

Also, its name is the same as the one with PanopticFCN-512 / R50 / 3x / faster threshold.

May I ask you to crosscheck the weight and its link?

Thanks!

yanwei-li commented 3 years ago

Hi, thanks for your interest. Actually, PanopticFCN-400 share the same model with that in PanopticFCN-512. That means it is trained with 512 size and test with 400 size, which would bring a better performance according to our experiments. And the faster version share the same model with the normal but with different inference setting, as declared in the README.

emjay73 commented 3 years ago

Oh, I get it. Thanks!