"""
Gaze direction estimation from MediaPipe Face Mesh iris landmarks.

Requires mediapipe to be run with refine_landmarks=True (enables iris tracking,
landmark indices 468-477).
"""

from __future__ import annotations

from dataclasses import dataclass

import numpy as np

_LEFT_IRIS_CENTER = 468
_RIGHT_IRIS_CENTER = 473
_LEFT_EYE_INNER = 133
_RIGHT_EYE_OUTER = 263


@dataclass(frozen=True)
class GazeDirection:
    dx: float  # [-1, 1] — negative = left, positive = right
    dy: float  # [-1, 1] — negative = up, positive = down

    @property
    def label(self) -> str:
        if abs(self.dx) < 0.15 and abs(self.dy) < 0.15:
            return "center"
        if abs(self.dx) > abs(self.dy):
            return "left" if self.dx < 0 else "right"
        return "up" if self.dy < 0 else "down"


class GazeEstimator:
    """Estimate gaze direction from iris center relative to eye corners."""

    def estimate(self, face_landmarks: np.ndarray) -> GazeDirection:
        """
        Args:
            face_landmarks: (478, 3) float32 — MediaPipe Face Mesh with iris refinement.

        Returns:
            GazeDirection with normalized (dx, dy).
        """
        left_iris = face_landmarks[_LEFT_IRIS_CENTER]
        right_iris = face_landmarks[_RIGHT_IRIS_CENTER]
        iris_center = (left_iris + right_iris) / 2.0

        left_inner = face_landmarks[_LEFT_EYE_INNER]
        right_outer = face_landmarks[_RIGHT_EYE_OUTER]
        eye_width = np.linalg.norm(right_outer - left_inner)
        if eye_width < 1e-6:
            return GazeDirection(dx=0.0, dy=0.0)

        eye_center = (left_inner + right_outer) / 2.0
        delta = iris_center - eye_center
        return GazeDirection(
            dx=float(delta[0] / eye_width),
            dy=float(delta[1] / eye_width),
        )