Xunjiang (Alfred) Gu

An ugly workaround is to set every torch inverse to follows: `torch.inverse(lidar2ego_rots.to("cpu")).to("cuda:0")`

The performance seems far off from what is being published. The Shapely version is 1.8.5.post1.

As you can see, the result is quite poor with the uploaded checkpoint on the nuscenes mini. I am wondering why this is the case?

@colahe On my local machine

@colahe Sorry for the late reply. It's been a while since I explored this repo. I don't quite remember the specific arguments but I think it is a similar command...

I do not quite remember, did you try on the entire val set? You can try visualizing it to see whether it is really poor or not