src.evaluation_metric.pck

src.evaluation_metric.pck(x: torch.Tensor, x_gt: torch.Tensor, perm_mat: torch.Tensor, dist_threshs: torch.Tensor, ns: torch.Tensor) torch.Tensor[source]

Percentage of Correct Keypoints (PCK) evaluation metric.

If the distance between predicted keypoint and the ground truth keypoint is smaller than a given threshold, than it is regraded as a correct matching.

This is the evaluation metric used by “Zanfir et al. Deep Learning of Graph Matching. CVPR 2018.”

Parameters
  • x\((b\times n \times 2)\) candidate coordinates. \(n\): number of nodes in input graph

  • x_gt\((b\times n_{gt} \times 2)\) ground truth coordinates. \(n_{gt}\): number of nodes in ground truth graph

  • perm_mat\((b\times n \times n_{gt})\) permutation matrix or doubly-stochastic matrix indicating node-to-node correspondence

  • dist_threshs\((b\times m)\) a tensor contains thresholds in pixel. \(m\): number of thresholds for each batch

  • ns\((b)\) number of exact pairs. We support batched instances with different number of nodes, and ns is required to specify the exact number of nodes of each instance in the batch.

Returns

\((m)\) the PCK values of this batch

Note

An example of dist_threshs for 4 batches and 2 thresholds:

[[10, 20],
 [10, 20],
 [10, 20],
 [10, 20]]