Implicit train-free calibration for video-based eye-tracking

Appearance-based gaze estimation methods based on deep learning perform significantly better when they have been appropriately calibrated at a per-participant (subject) level. However, their calibration process typically includes neural model retraining with ground truth subject gaze data, which is difficult to obtain, leaving much room for error and consuming a non-negligible portion of the recording time. To address this issue, we propose a novel train-free calibration scheme, which includes a novel neural architecture and training process that learns to operate with implicit calibration, by design. More specifically, the input image representation is refined by extracting information about the visual similarity between the input image and the proposed calibration anchors, i.e., representative images of subjects linked with rough gaze directions, using an attention mechanism. During deployment, the model is adapted to new subjects by enriching the input image representation with its similarity to a set of representative test subject images, without model retraining or ground truth gaze data. Our experiments in publicly available eye-tracking datasets have shown that the proposed method provides almost a 10-15% reduction in angular error with respect to baseline solutions.