请问occupancy有大model的config么？ about tpvformer HOT 3 OPEN

Sorry for the confusion. We actually use [100*100*8 tpv resolution, 256 feature dimension] for 3D semantic occupancy prediction and [200*200*16 tpv resolution, 128 feature dimension] for lidar segmentation in the paper. Note that there is no necessary connection between the resolution of the tpv planes and the voxel resolution for visualization, since we can upsample the tpvplanes as shown in Fig. 6 at test time.
On the other hand, finer details could be expected if tpv planes of higher resolution are used.
Also, thanks for reporting the bug to us.

from tpvformer.

I see. Thanks for the reply.
I can understand we can upsample the tpvplanes during test time. But is there any reason not to use 200x200x16 for training in your paper? Did you observe minor improvements when increasing from 100x100x8 (with 2x upsample at test time) to 200x200x16 (no upsample at time time)?

from tpvformer.

In fact, we did not notice substantial improvement qualitatively through visualization, when training with a resolution of 200x200x16. We think it might be due to the sparse nature of LiDAR supervision, which is further sparsified with higher resolution.

from tpvformer.