- How many Observations are Enough? Knowledge Distillation for Trajectory Forecasting
Motivation: Motivation is the most important part in this paper. Existing protocol 8-12 is not partical, because "the input trajectories required to condition the prediction have to be automatically extracted by a tracking system" (i.e., noisy, human intervention, etc.).
Method: Propose a knowledge distillation-based approach, backbone adopted from Transformer.
My thoughts: