The mIoU of Uniocc and RenderOcc is not same

[HenryZhangJianhe]

Uniocc CVPR2023-3D-Occupancy-Prediction nuscenes is 51.27 RenderOcc occ3d-nuscens 26.11 是因为这两个测试数据集的gt 差别很大导致性能差别很大？

[pmj110119]

The former score is based on the use of camera_mask during training, which can bring huge improvement to the metric, but the visualization result will become worse. This is explained in Section 2.4 of UniOcc.

In the CVPR2023 Challenge that UniOcc participated in, camera_mask was allowed to be used by the organizer. However, the use of this mask is controversial for research, and I recommend following the Occ3D paper and not using it.

[99er-gao]

@pmj110119 Hello author, thank you for your excellent work. In Section 2.4, it is mentioned that "employee mask camera for loss during training". May I ask if mask_camera is only used in loss_3d? If it is also used in loss_rendering, how is it utilized? I am very confused and looking forward to your reply. Thank you!

[pmj110119]

@pmj110119 Hello author, thank you for your excellent work. In Section 2.4, it is mentioned that "employee mask camera for loss during training". May I ask if mask_camera is only used in loss_3d? If it is also used in loss_rendering, how is it utilized? I am very confused and looking forward to your reply. Thank you!

Yes, mask_camera can only be used for loss_3d.