src.feature_align.interp_2d

src.feature_align.interp_2d(z: torch.Tensor, P: torch.Tensor, ori_size: torch.Tensor, feat_size: torch.Tensor, out=None, device=None) → torch.Tensor[source]

Interpolate in 2d grid space. z can be 3-dimensional where the first dimension is feature dimension.

Parameters

  • z – \((c\times w\times h)\) feature map. \(c\): number of feature channels, \(w\): feature map width, \(h\): feature map height

  • P – \((n\times 2)\) point set containing point coordinates. The coordinates are at the scale of the original image size. \(n\): number of points

  • ori_size – \((2)\) size of the original image

  • feat_size – \((2)\) size of the feature map

  • out – optional output tensor

  • device – output device. If not specified, it will be the same as the input

Returns

\((c \times n)\) extracted feature vectors