[✨ FEATURE] : Auto model select option for gaze estimation and calibration

[sohampirale]

Summary

Currently, the regression model used for gaze estimation is manually selected by the UX researcher/admin from the frontend and passed to the backend for training during calibration.

This requires users to understand the differences between available regression models, which may not always be realistic or optimal.

This feature proposes adding an “Auto” model selection option that automatically determines the best-performing model during calibration using validation error metrics (currently computed as RMSE per axis).

---

Problem

• Model selection is manual and exposed to non-ML users.
• There is no guarantee that the selected model produces the lowest validation error for a given user.
• Calibration performance may vary across users due to lighting, device setup, camera position, and facial geometry.
• When multiple calibration documents exist for a user, there is no systematic way to determine which calibration/model combination performs best.

---

Proposed Solution

Introduce an “Auto” model selection option with the following behavior:

1. During Calibration

• Use training phase data to train all candidate regression models.
• Use validation split (currently 80/20) to evaluate each trained model.
• Compute validation performance using the existing RMSE-based metric (currently labeled as “accuracy”).
• Aggregate validation error across all calibration points.
• Select the model with the lowest validation error (RMSE).

2. Persist Calibration Metadata

Store the following in the calibration document:

• Selected model name
• Validation error (RMSE)
• Per-model validation metrics (optional but recommended)
• Timestamp and calibration metadata

3. During Gaze Analysis

• Use the selected model associated with that calibration.
• If a user has multiple calibration documents, allow selecting or automatically reusing the calibration with the lowest stored validation error.

---

Benefits

• Removes ML decision burden from UX researchers.
• Ensures per-user model optimization using objective validation metrics.
• Enables data-driven reuse of the most accurate calibration.

---

Would appreciate feedback on this approach

[sohampirale]

I am working on this

[sohampirale]

• I'm also thinking about a unified evaluation approach using spatial accuracy (Euclidean distance error) to fairly
  compare all models regardless of their training method. This would address the inconsistency where some models are optimized with R² scores while others use default parameters, allowing for true apples-to-apples comparison based on the metric that matters most for eye tracking accuracy.

• The diagram illustrates how all 8 models would be evaluated with the same spatial accuracy metric post-training, enabling selection of the model with the lowest Euclidean error regardless of whether it used hyperparameter tuning or default parameters.

• Would appreciate feedback from the maintainers on whether this unified evaluation approach aligns

@marcgc21

[sohampirale]

This auto-model selection feature would work well with a phased calibration approach:

Current Approach:

• Training Phase → Validation Phase → Send BOTH training + validation data in validation requests
• Model training happens during validation

Proposed Approach:

• Training Phase: Send training data → Backend trains and stores model
• Validation Phase: Send validation data → Backend reuses trained model
• Training happens once, validation reuses model

If this approach sounds good,then it involves changes across multiple repositories (RUXAILAB frontend, web-eye-tracker-front, and eye-tracker-api), I'd appreciate any feedback on this