how can I compute accuracy according to predicted location instead of heatmaps?

[zoeccolor]

how can I compute accuracy according to predicted location instead of heatmaps? And I don not understand the function calc_integral_accuracy in alphapose/utils/metrics.py, can u give me an example? thx

[nistring]

In my opinion, it might be look inappropriate to assess accuracy of locations of predicted keypoints or heatmaps. Instead, pose estimation adopts various metrics for assessment such as PDJ, PCK, average precision using OKS etc. The metric highly depends on the annotation style of dataset for example, average precision is applicable to COCO dataset while MP2 dataset chose PCKh. I suppose PCP(Percentage of Correct Keypoints) would suit your interest in that it excludes every prediction outside of certain distance,