ten31-transcripts/Ten31Transcripts/Visual/GridCallAnalyzer.swift

import Foundation
import CoreGraphics
import CoreVideo
import CoreImage

/// Shared engine for tile-grid conferencing UIs: OCR the name/initials on each
/// tile, then mark the active speaker(s) by the speaking-highlight around their
/// tile. Handles BOTH highlight kinds:
/// - **white border** (Signal: 3px #ffffff ring — detected via thin near-white edges)
/// - **coloured border** (Zoom/Teams — detected via saturated edges)
///
/// The white name text is excluded so it can't be mistaken for the white border.
/// Geometry (`Config`) is a first pass; tile expansion calibrates per app against
/// real screenshot fixtures. Detection *logic* is validated on synthetic frames.
struct GridCallAnalyzer {
    struct Config {
        var tileExpandX = 2.4         // tile width  ≈ name width  × this
        var tileExpandY = 4.8         // tile height ≈ name height × this
        var nameAtBottom = true       // Signal/most: name footer sits at the tile bottom
        var detectColoredBorder = true
        var detectWhiteBorder = true
        var minTextConfidence: Float = 0.3
        var maxNameLength = 40
        var minHighlightPoints = 6
        var highlightShareOfMax = 0.35
    }

    var config = Config()
    var recognizer = TextRecognizer()

    func analyze(pixelBuffer: CVPixelBuffer, at t: TimeInterval) -> [SpeakerObservation] {
        guard let cg = Self.cgImage(from: pixelBuffer) else { return [] }
        return analyze(cgImage: cg, at: t)
    }

    func analyze(cgImage: CGImage, at t: TimeInterval) -> [SpeakerObservation] {
        let texts = recognizer.recognize(in: cgImage).filter {
            $0.confidence >= config.minTextConfidence && !cleaned($0.text).isEmpty
        }
        guard !texts.isEmpty, let sampler = FrameSampler(cgImage: cgImage) else { return [] }
        let w = cgImage.width, h = cgImage.height

        struct Tile { let name: String; let textRect: CGRect; let tile: CGRect; let conf: Double }
        let tiles = texts.map { r in
            Tile(name: cleaned(r.text),
                 textRect: pixelRect(r.boundingBox, w, h),
                 tile: tileRect(r.boundingBox, w, h),
                 conf: Double(r.confidence))
        }

        // Highlight pixels: coloured (saturated) and/or white (thin near-white).
        var highlight: [CGPoint] = []
        if config.detectColoredBorder { highlight += sampler.saturatedPoints() }
        if config.detectWhiteBorder { highlight += sampler.thinWhitePoints() }

        // Drop points inside any name-text region so the white name itself doesn't count.
        let exclusions = tiles.map {
            $0.textRect.insetBy(dx: -$0.textRect.width * 0.15, dy: -$0.textRect.height * 0.35)
        }
        let points = highlight.filter { p in !exclusions.contains { $0.contains(p) } }

        // Attribute each highlight pixel to EXACTLY ONE tile — the (no-margin)
        // estimated rect that contains it, nearest centre as tiebreak. Containment
        // (not a radius) keeps a border from bleeding into adjacent tiles even when
        // the tile-size estimate is rough; an under-sized estimate merely drops the
        // far edge rather than misattributing it.
        var counts = [Int](repeating: 0, count: tiles.count)
        for p in points {
            var best = -1
            var bestDistSq = Double.greatestFiniteMagnitude
            for (i, tile) in tiles.enumerated() where tile.tile.contains(p) {
                let dx = p.x - tile.tile.midX, dy = p.y - tile.tile.midY
                let dd = dx * dx + dy * dy
                if dd < bestDistSq { bestDistSq = dd; best = i }
            }
            if best >= 0 { counts[best] += 1 }
        }
        let maxCount = counts.max() ?? 0
        let need = max(config.minHighlightPoints, Int(Double(maxCount) * config.highlightShareOfMax))

        return tiles.enumerated().map { idx, tile in
            let speaking = maxCount >= config.minHighlightPoints && counts[idx] >= need
            return SpeakerObservation(name: tile.name, speaking: speaking,
                                      bbox: tile.tile, confidence: tile.conf, t: t)
        }
    }

    /// Vision normalized bbox (bottom-left origin) → pixel rect (top-left origin).
    private func pixelRect(_ box: CGRect, _ w: Int, _ h: Int) -> CGRect {
        let W = Double(w), H = Double(h)
        return CGRect(x: box.minX * W, y: (1 - box.maxY) * H, width: box.width * W, height: box.height * H)
    }

    /// Estimate the participant tile from the name label. With `nameAtBottom`, the
    /// tile extends UP from the footer (Signal); otherwise it's centred on the name.
    private func tileRect(_ box: CGRect, _ w: Int, _ h: Int) -> CGRect {
        let W = Double(w), H = Double(h)
        let name = pixelRect(box, w, h)
        let nw = name.width * config.tileExpandX
        let nh = name.height * config.tileExpandY
        let cx = name.midX
        let rect: CGRect
        if config.nameAtBottom {
            let bottom = name.maxY + name.height * 0.3
            rect = CGRect(x: cx - nw / 2, y: bottom - nh, width: nw, height: nh)
        } else {
            rect = CGRect(x: cx - nw / 2, y: name.midY - nh / 2, width: nw, height: nh)
        }
        return rect.intersection(CGRect(x: 0, y: 0, width: W, height: H))
    }

    private func cleaned(_ s: String) -> String {
        let t = s.trimmingCharacters(in: .whitespacesAndNewlines)
        return t.count <= config.maxNameLength ? t : ""
    }

    private static let ciContext = CIContext()

    static func cgImage(from pixelBuffer: CVPixelBuffer) -> CGImage? {
        let ci = CIImage(cvPixelBuffer: pixelBuffer)
        return ciContext.createCGImage(ci, from: ci.extent)   // reuse; allocating per frame is costly
    }
}