code on 2D backbone about cylinder3d HOT 6 CLOSED

When implementing the 2D backbone, our cylinder3D performs a 2D transformation following PointPillar, from 3D space to bird-view representation, and then use the 2D conv to stack the 2D backbone. For asymmetrical block and DDCM, they are all implemented in 2D space. BTW, please refer to our latest version of Cylinder3D

from cylinder3d.

Thanks for your explanation. Is there any plan to release the code of 2D backbone? I'm trying to deploy this model onboard but unfortunately spconv is not convertible with TensorRT, so I would have to employ the 2D backbone.

from cylinder3d.

The key difference is the 3D-2D transformation, and you can follow this code to function it.

from cylinder3d.

I'm familiar with the PointPillar 3D-2D transformation. Could you elaborate more on what would asymmetrical block and DDCM look like in 2D?

from cylinder3d.

For asymmetrical block, its shape looks like 3x1 and 1x3 and no third dimension. For DDCM, you can refer to this paper.

from cylinder3d.

That's super helpful, thanks!

from cylinder3d.