Anatomical Domain Shifts: Test-time Heterogeneous Adaptation for 3D Human Pose Prediction

Figure 1. Proof of Concept: Anatomical Domain Shift. t-SNE embedding of human segments based on anatomical separation, i.e., left/right legs, left/right arms, and torso, and full body under the challenging domain shift setup for TTA-based HPP [11]. This plot shows the feature topology of the source domain-H3.6M [24] (upper), and the target domain-GRAB [48] (bottom). The diagram exhibits that while full-body embedding differ between source and target domains, such matter is not present on all segments, and as a contrast, the distribution of certain segments (i.e., right leg and torso) is very close. This provides the proof of concept of our central hypothesis: for human motion, domain shift manifests primarily at human anatomical segments rather than the full-body.