This module contains implementation of autograd operations performed directly on point cloud.

class d3d.point.AlignedScatter[source]

Bases: torch.autograd.function.Function

d3d.point.aligned_scatter(coordinates, feature_map, method='drop')[source]

Gather the values given coordinates in feature_map.

  • feature_map – should have shape B x C x D1 x D2… x Dm, the number of remaining dimensions (m) should be consistent with coordinate dimension

  • coordinates – should have shape N x (m+1), it should contains batch index at the first dimension

  • method – drop - directly convert decimal coordinates to integers mean - mean of surrounding values linear - (bi)linear interpolation of surrounding values max - maximum of surrounding values nearest - value of nearest value


extracted features with shape N x C

Note: right now only drop, mean and linear are implemented