Increase the min calibration points to 7 instead of 2 to avoid overfitting

[midaa1]

Description

This PR introduces a minimum threshold for calibration points to ensure that the gaze prediction models are trained on a sufficient number of samples.

Previously, calibration could proceed with too few points, leading to unstable models and unreliable evaluation results.

Problem

• Calibration with very few points can result in overfitting and noisy predictions.

• Model performance becomes highly sensitive to outliers when the sample size is too small.

• Evaluation metrics are not meaningful with insufficient calibration data.

What Changed

• Enforced a minimum of 7 calibration points before training is allowed.

• Improved overall robustness and consistency of gaze prediction results by increasing the number of samples per point to 100.

Rule of Thumb

As a general rule of thumb in regression-based calibration tasks:

A minimum of 5–10 well-distributed calibration points is required to obtain a stable mapping.

Setting the minimum to 7 points provides a reasonable balance between usability and statistical reliability.

This ensures:

Better generalization

More stable model coefficients

More meaningful evaluation metrics

[midaa1]

@marcgc21 @jvJUCA check this tiny change to avoid overfitting