pillar-based voxelization, and the resulting pseudo image is in BEV, and as an input to a 2D-like Siamese SOT method.
Objects in the Bird’s-eye View (BEV) coordinates, have constant size. Thus, only the object rotation can change in the new coordinate system and not the object scale. For this reason, we replace multi-scale search with a multi-rotation search.
Why? different devices different success ratio?