"""Neuro-inspired scoring for a detected scene.

These scores are a SIMULATION for visualization, not a neuroscience measurement
and not a medical diagnosis. They are simple, transparent heuristics derived from
the object detections.
"""

from statistics import fmean, pvariance


def _box_center(box: dict) -> tuple:
    return (
        (box["xmin"] + box["xmax"]) / 2.0,
        (box["ymin"] + box["ymax"]) / 2.0,
    )


def compute_scores(detection: dict) -> dict:
    """Build 0-100 scores from a detection result (see detection.detect)."""
    objects = detection["objects"]
    size = detection["image_size"]
    width = max(size["width"], 1)
    height = max(size["height"], 1)

    object_count = len(objects)

    # More objects -> more clutter.
    visual_clutter = min(object_count * 12, 100)

    # Spatial spread: how scattered the object centers are across the frame.
    # Normalized variance of centers (0 = all stacked, ~100 = spread to corners).
    if object_count >= 2:
        cx = [_box_center(o["box"])[0] / width for o in objects]
        cy = [_box_center(o["box"])[1] / height for o in objects]
        spread = (pvariance(cx) + pvariance(cy))  # each variance in [0, 0.25]
        spread_score = min(spread / 0.5 * 100, 100)
    else:
        spread_score = 0.0

    # Harder to focus when there are many objects AND they are spread out.
    focus_difficulty = round(0.6 * visual_clutter + 0.4 * spread_score)

    # Confidence of the single most prominent object.
    if objects:
        top = objects[0]
        primary_object = {
            "label": top["label"],
            "score": round(top["score"] * 100),
        }
    else:
        primary_object = {"label": None, "score": 0}

    scene_complexity = round(fmean([visual_clutter, focus_difficulty]))

    return {
        "object_count": object_count,
        "visual_clutter": int(visual_clutter),
        "focus_difficulty": int(focus_difficulty),
        "primary_object": primary_object,
        "scene_complexity": int(scene_complexity),
    }