GazeEval
The first computational method for predicting the perceived visual realism of AI-generated videos from eye-tracking data; it takes eye-movement sequences (fixation coordinates, durations, saccade paths) as input and outputs the tendency for a video to be perceived as real.
Timeline (last 90 days)
GazeEval achieves 71.3% accuracy on the prediction task.
Possible application scenarios for methods like GazeEval include helping locate the specific regions in generated videos that trigger a sense of "unrealness," providing perceptual-level feedback signals for model training, and adding an automatable quality screening step beyond human evaluation.
All Facts (2)
GazeEval achieves 71.3% accuracy on the prediction task.
50%UnverifiedPossible application scenarios for methods like GazeEval include helping locate the specific regions in generated videos that trigger a sense of "unrealness," providing perceptual-level feedback signals for model training, and adding an automatable quality screening step beyond human evaluation.
50%