What is meant by features, and coords!?

[a-akram-98]

In this code:

class Voxelization(nn.Module):
    def __init__(self, resolution, normalize=True, eps=0):
        super().__init__()
        self.r = int(resolution)
        self.normalize = normalize
        self.eps = eps

    def forward(self, features, coords):
        coords = coords.detach()
        norm_coords = coords - coords.mean(2, keepdim=True)
        if self.normalize:
            norm_coords = norm_coords / (norm_coords.norm(dim=1, keepdim=True).max(dim=2, keepdim=True).values * 2.0 + self.eps) + 0.5
        else:
            norm_coords = (norm_coords + 1) / 2.0
        norm_coords = torch.clamp(norm_coords * self.r, 0, self.r - 1)
        vox_coords = torch.round(norm_coords).to(torch.int32)
        return F.avg_voxelize(features, vox_coords, self.r), norm_coords

    def extra_repr(self):
        return 'resolution={}{}'.format(self.r, ', normalized eps = {}'.format(self.eps) if self.normalize else '')

What is the input Features and Coords, In Kitti Dataset for example point cloud are defined as (N,4) as 3 for x,y,z, and one for reflectance so what is the features and coords from those 4 and why coords.detach() this will cut the gradient flow !?

[zhijian-liu]

coords will be [x, y, z] and feats will be [x, y, z, reflectance]. We apply the detach operation to coords because its gradient is not required (and it Is not differentiable anyway after the torch.round operation).

[zhijian-liu]

I'm closing this issue due to inactivity. Please feel free to reopen it if the problem has not been resolved.