Figure 10. Pose correction using pictorial structures.

(A) Raw input data from four cameras, focusing on the pretarsus of the middle left leg. (B) Probability maps for the pretarsus output from the Stacked Hourglass deep network. Two maxima (white arrowheads) are present on the probability maps for camera 5. The false-positive has a larger unary probability. (C) Raw predictions of 2D pose estimation without using pictorial structures. The pretarsus label is incorrectly applied (white arrowhead) in camera 5. By contrast, cameras 4, 6, and 7 are correctly labeled. (D) Corrected pose estimation using pictorial structures. The false-positive is removed due to the high error measured in Equation 8. The newly corrected pretarsus label for camera five is shown (white arrowhead).