Human Pose Matching

Abstract

We present a novel approach to accurately compare poses of multiple persons in 2D pictures. The algorithm compares a model picture with an input picture and returns a score on how well they match disregarding the physiques of the persons. The poses on the pictures are defined by a pose estimator for both the model and input picture. A pose is composed by key points. Using these key points an affine transformation is computed. This searches for the best fit between the key points with only performing a translation, rotation, shear and scaling on the key point set. All these transformations don’t affect the pose except for the rotation factor. This factor is used in the calculation of the score as well as the euclidean distance between the transformed input and the original model key points. The algorithm also compares the spatial relation between multiple persons on picture. To handle multiple persons in a picture, all the poses are correlated individually. To compare the interaction between poses, an affine transformation is again calculated with the key points of all poses.