Hi,
in the README you say that DS-Net is based on the original version of Cylinder3D and link to the paper that reports 64.3% mIoU on the SemanticKITTI val set.
I compared the layers in the pretrained weights you provide here with the pretrained weights of the current Cylinder3D version that are provided in their repo.
It seems that the only difference is that MODEL.VFE.OUT_CHANNEL = 64 in the "original" version and 256 in the current version.
With this parameter change I evaluated the "sem_pretrain.pth" on the SemanticKITTI val set and got an mIoU of 58.5% which is very far away from the 64.3% reported in the paper.
Can you explain how you trained the pretrained model of step 1 and what exactly is the difference to the latest version of Cylinder3D?
Thank you
Hi,
in the README you say that DS-Net is based on the original version of Cylinder3D and link to the paper that reports 64.3% mIoU on the SemanticKITTI val set.
I compared the layers in the pretrained weights you provide here with the pretrained weights of the current Cylinder3D version that are provided in their repo.
It seems that the only difference is that MODEL.VFE.OUT_CHANNEL = 64 in the "original" version and 256 in the current version.
With this parameter change I evaluated the "sem_pretrain.pth" on the SemanticKITTI val set and got an mIoU of 58.5% which is very far away from the 64.3% reported in the paper.
Can you explain how you trained the pretrained model of step 1 and what exactly is the difference to the latest version of Cylinder3D?
Thank you