huge long time to generate trainval GT

[qihao-plus]

Thanks for your work.

But I find that it takes huge long time to generate trainval GT even with parallel generating. Also, it takes almost all CPU threads.

Is that same with your past practice?

Thanks!

[FANG-MING]

Yes, non-keyframe point clouds lack labels, requiring the use of nearest neighbor methods to assign labels to unlabeled points, which can be a time-consuming task.

[qihao-plus]

Thanks for your reply. Yes, it takes 96 CPU servers days. BTW, what’s the difference with CVPR 2023 occupancy GT ? Is that the same method to generate GT as yours?

Also, the recent OpenOccupancy repo releases it’s code as well for nuscenes. What’s the difference with your GT?

Thanks!

[FANG-MING]

Yes, this speed is normal. The specific content of OpenOccupancy can refer to their papers.

[SISTMrL]

hello, could you please release the generated occupancy data?