Closed jtengyp closed 10 months ago
Hi, @jtengyp. In our implementation, for uniad-small with 100x100 bev size, we actually train the perception heads(track_map) on top of the 100x100 bev plane, then the bev feature is upsampled to 200x200 for prediction and planning heads. That's why the upsample_bev_if_tiny function is needed. We will release the configs for uniad-small soon, which is an example of how to adapt to smaller bev.
@YTEP-ZHI Will the configs for uniad-small be released?
Hi @jtengyp, as we have other projects going on, we may not be available to release other UniAD variants in the near future.
The configs only provide bev size of 200x200. I'm trying difference bev size, like 100x100, or 50x50.
I noticed that in NuScenesE2EDataset, the canvas_size only supports 50 and 200.
While in UniADTrack->upsample_bev_if_tiny, the bev_size only supports 100x100.
Now I'm confused about bev size of 50 and 100. Or the bev size of 50 does not need to upsample in UniADTrack->upsample_bev_if_tiny?