Transferring Dense Pose to Proximal Animal Classes.

摘要: Recent contributions have demonstrated that it is possible to recognize the pose of humans densely and accurately given a large dataset poses annotated in detail. In principle, same approach could be extended any animal class, but effort required for collecting new annotations each case makes this strategy impractical, despite important applications natural conservation, science business. We show that, at least proximal classes such as chimpanzees, transfer knowledge existing dense recognition humans, well more general object detectors segmenters, problem other classes. do by (1) establishing DensePose model which also geometrically aligned (2) introducing multi-head R-CNN architecture facilitates multiple tasks between classes, (3) finding combination known can transferred most effectively (4) using self-calibrated uncertainty heads generate pseudo-labels graded quality training class. introduce two benchmark datasets labelled manner class chimpanzee use them evaluate our approach, showing excellent learning performance.