Abstract
To successfully track the puck when watching an ice hockey game the oculomotor system relies on sensory input that is lagging behind its correct position due to processing delays and motor latencies. To compensate for these delays it is necessary to make predictions about the correct position of the puck, which can get very difficult for complex movements with rapid changes in direction. We investigated whether context information (for example the player movements) contributes to these predictions by showing participants short clips of ice hockey games (10s, stationary camera) while tracking their eyes. Participants either saw the regular clip (context) or a still image of the first frame with a black dot moving along the hand-labeled puck positions (no-context). Cross-correlation analysis demonstrated that the peak correlation between eye and target movement was present with around 200 ms delay in the no-context condition, whereas the delay was significantly reduced to around 50 ms when context was available. For passes between players we observed that participants used the context information and produced predictive saccades to receiving players around 300 ms before the puck arrived. Additionally, also stopping the eye at the target location of the pass was more accurate with context, as participants significantly overshot the location of the end of the pass in the no-context condition. To gain insight into the underlying computations, we compared eye movements in the context condition to the output of a deep network model trained to use the imagery and optic flow to predict puck position. Interestingly, after partialing out the ground truth puck positions, gaze and network were still correlated, suggesting the network and the observers use similar context cues to follow the puck. Overall, our results show that the oculomotor system can efficiently use context information to counter internal processing delays by making predictions.