Figure 1. Deriving 3D pose from multiple camera views.

(A) Raw image inputs to the Stacked Hourglass deep network. (B) Probability maps output from the trained deep network. For visualization purposes, multiple probability maps have been overlaid for each camera view. (C) 2D pose estimates from the Stacked Hourglass deep network after applying pictorial structures and multi-view algorithms. (D) 3D pose derived from combining multiple camera views. For visualization purposes, 3D pose has been projected onto the original 2D camera perspectives. (E) 3D pose rendered in 3D coordinates. Immobile thorax-coxa joints and antennal joints have been removed for clarity.

Figure 1—video 1. Deriving 3D pose from multiple camera views during backward walking in an optogenetically stimulated MDN>CsChrimson fly.

Download video file ^{(11.2MB, mp4)}

DOI: 10.7554/eLife.48571.003