PointsCoder/VOTR

Dilated Attention

Yzichen opened this issue · 2 comments

I don’t quite understand the implementation of Dilated Attention and the setting of RANGE_SPEC.
If I want to get the result of Fig3(2D example) in the paper, how to set the parameters.

Screenshot from 2021-09-28 16-26-12

Take the 3rd item in RANGE_SPEC as an example:

[5, 25, 5, 5, 25, 5, 0, 15, 2] is [xstart, xend, xstride, ystart, yend, ystride, zstart, zend, zstride], which means the searching range for x is [5, 10, 15, 20], for y is [5, 10, 15, 20], for z in [0, 2, 4, 6, 8, 10, 12, 14], so querying indices are the combination of xyz range, e.g. (5, 20, 14).

Each item in RANGE_SPEC contains one searching range. The searching process will end until the number of founded non-empty voxels reaches SIZE (48).

If you want to visualize the neighboring voxels, you can choose to visualize the variable key_coords in votr_backbone.py

Closed if no other questions.