The task requires drawing visual figures that share underlying structure, based onasetofcomposable geometric rulesandsimpleobjects.

However,thesparsityofaccount activities onsocial media limits theperformance ofexisting deep learning based coordination detectors as they can not exploit useful prior knowledge.

Videos introduce intrinsic complexities inboth spatial and temporal dimensions, whichare not present in staticimages.

Figure 14: SpeedupofTheseus(forwardpass) over Ceres (black dashed) ondifferent PGOproblemscales. Fig. (right) degradation, right) is numerical priorwork.