search

hongfz16 / DS-Net

[CVPR 2021/TPAMI 2023] Rank 1st in the public leaderboard of SemanticKITTI Panoptic Segmentation (2020-11-16)
MIT License
243 stars 29 forks source link

Difference between original and latest version of Cylinder3D #12

Open heumchri opened 3 years ago

heumchri commented 3 years ago

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

hongfz16 commented 2 years ago

Hi, please checkout the Cylinder3D repo for the training details. The latest version of Cylinder3D has many differences from that used in this repo other than the channel difference. Please kindly check the CVPR paper of Cylinder3D (Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR Segmentation) for the network details, which may result in this drastic performance drop. Thanks.