Haiyang-W / DSVT

[CVPR2023] Official Implementation of "DSVT: Dynamic Sparse Voxel Transformer with Rotated Sets"
https://arxiv.org/abs/2301.06051
Apache License 2.0
353 stars 28 forks source link

Question about the setting of voxel version #67

Closed 3bobo closed 6 months ago

3bobo commented 6 months ago

Thanks for your great work! However, the setting of the voxel version confused me. https://github.com/Haiyang-W/DSVT/blob/a0ab685b573896fc4bc9a64b187bdc2b574c3741/tools/cfgs/dsvt_models/dsvt_3D_1f_onestage.yaml#L78

Each stage only has 1 block. Therefore, for each stage, it can only use non-shift or shift, not both. For example, in stage-0, the block only uses the window [12, 12, 32] with shift [0, 0, 0] X and Y set_voxel_inds. But in stage-1, the block only uses the window [24, 24 8] with shift [6, 6, 0] X and Y set_voxel_inds. Therefore, within each stage, is there no shift and hybrid setting to propagate the inter-window feature?

chenshi3 commented 6 months ago

Yes, there is no shift and hybrid setting within one stage. However, shift window between stages can still provide global receptive field.

3bobo commented 6 months ago

Thanks for your reply!