Figure 5
Handling ambiguity. This observer AI is trained only on the first frame as input. Therefore, there are multiple possible future trajectories possible based on only this input frame. As shown in (A). Image Pair Prediction, the observer handles this ambiguity by outputting one of the heuristic behaviors or as a blurry trajectory. Although our observer model is trained using start and end image pairs (with integration of all the past frames), the observer can be used in an online fashion during evaluation. (B) Progressive Video Sequence Prediction shows the prediction results using our model across multiple time stamps.
