mit-han-lab / spvnas

[ECCV 2020] Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution
http://spvnas.mit.edu/
MIT License
585 stars 109 forks source link

Question on devoxelization #102

Closed junshengzhou closed 2 years ago

junshengzhou commented 2 years ago

Hi, Thanks for your great job! I have one question on the devoxelization process. I found that the coordinates of raw points is lost by applying

pc_ = np.round(block[:, :3] / self.voxel_size).astype(np.int32) 
pc_ -= pc_.min(0, keepdims=1)

to get the voxelized coordinates. And only the voxelized coordinates are passed to the voxelization and devoxelization processes. So I wonder how to do trilinear interpolation for get point-specific features in devoxelization without their actual coordinates? Or are the features of different points that located in one voxel the same?

zhijian-liu commented 2 years ago

Yes, the coordinates of raw points are lost. Please note that the point locations for trilinear interpolation are also quantized.

zhijian-liu commented 2 years ago

I'm closing this issue due to its inactivity. However, please feel free to reopen it if your problem persists.