Closed adammpolak closed 2 years ago
We find such 3D translation vector (the root point coordinates), which gives minimum of projection error from 3D coordinates (predicted as root-relative shifts) to 2D predicted keypoints (check the paper for the derivation). Each human is assumed to be of the same height, 180 cm.
Hope, it helped.
@Daniil-Osokin
Would I be able to increase the minimum of projection error and then be able to get detections at further distances?
I read through the paper and I cannot find any metrics on how far away a person can be and still be detected using this model.
Is there a pixel height of a person that is required by the network or a general distance?