Intention-based Long-Term Human Motion Anticipation

Recently, a few works have been proposed to model theuncertainty of the future human motion. These works donot forecast a single sequence but multiple sequences forthe same observation. While these works focused on in-creasing the diversity, this work focuses on keeping a highquality of the forecast sequences even for very long timehorizons of up to 30 seconds. In order to achieve this goal,we propose to forecast the intention of the person ahead oftime. This has the advantage that the generated human mo-tion remains goal oriented and that the motion transitionsbetween two actions are smooth and highly realistic. Wefurthermore propose a new quality score for evaluation thatcorrelates better with human perception than other metrics.The results and a user study show that our approach fore-casts multiple sequences that are more plausible comparedto the state-of-the-art.