File size: 8,371 Bytes
4635d45
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
// MediaPipe Face Landmarker wrapper (browser only) β€” the web counterpart of
// app/face_tracker.py. Loads @mediapipe/tasks-vision from CDN; detect() returns
// blendshapes + head euler angles + normalized landmarks, or null when no face.
import { CDN_URL, WASM_BASE, MODEL_URL, eulerFromMatrix } from "./trackconfig.js?v=20260729.6";

// Load MediaPipe only when tracking or image auto-detection is requested. A
// static CDN import prevents the entire authoring UI from starting when the
// user is offline or the CDN is temporarily unavailable.
let tasksVisionPromise = null;
let tasksVisionAttempt = 0;
async function loadTasksVision() {
  if (!tasksVisionPromise) {
    // Browsers cache a rejected module import by URL. Give retries a distinct
    // URL so the retry button really performs another CDN request instead of
    // immediately replaying the earlier offline failure.
    const url = tasksVisionAttempt ? `${CDN_URL}?retry=${tasksVisionAttempt}` : CDN_URL;
    tasksVisionPromise = import(url).catch((err) => {
      tasksVisionPromise = null; // a later retry may succeed after reconnecting
      tasksVisionAttempt += 1;
      throw new Error(`could not load MediaPipe tasks-vision from CDN (offline?): ${err.message}`);
    });
  }
  return tasksVisionPromise;
}

export async function createTracker() {
  let vision;
  let FaceLandmarker, FilesetResolver;
  try {
    ({ FaceLandmarker, FilesetResolver } = await loadTasksVision());
    vision = await FilesetResolver.forVisionTasks(WASM_BASE);
  } catch (e) {
    throw new Error(e.message ?? `could not load MediaPipe tasks-vision wasm from CDN (offline?)`);
  }

  const options = {
    baseOptions: { modelAssetPath: MODEL_URL, delegate: "GPU" },
    runningMode: "VIDEO",
    numFaces: 1,
    outputFaceBlendshapes: true,
    outputFacialTransformationMatrixes: true,
  };

  let landmarker;
  try {
    landmarker = await FaceLandmarker.createFromOptions(vision, options);
  } catch (gpuErr) {
    try {
      options.baseOptions.delegate = "CPU"; // GPU delegate unavailable β€” retry on CPU
      landmarker = await FaceLandmarker.createFromOptions(vision, options);
    } catch (cpuErr) {
      throw new Error(`could not create FaceLandmarker (model download failed?): ${cpuErr.message}`);
    }
  }

  return {
    /** Detect one face in a <video>; return an observation or null when none. */
    detect(videoEl, tsMs) {
      if (!videoEl.videoWidth) return null; // not decoding yet
      const result = landmarker.detectForVideo(videoEl, tsMs);
      if (!result.faceBlendshapes || result.faceBlendshapes.length === 0) return null;

      const blend = {};
      for (const c of result.faceBlendshapes[0].categories) blend[c.categoryName] = c.score;

      // Raw landmark objects ({x, y, z}) β€” no per-frame array rebuild; consumers
      // read .x/.y directly (only the debug overlay uses these).
      const landmarks = result.faceLandmarks?.[0] ?? null;

      let yaw = 0, pitch = 0, roll = 0;
      if (result.facialTransformationMatrixes && result.facialTransformationMatrixes.length) {
        ({ yaw, pitch, roll } = eulerFromMatrix(result.facialTransformationMatrixes[0].data));
      }
      return { blend, yaw, pitch, roll, landmarks };
    },
    close() {
      landmarker.close();
    },
  };
}

// Worker init that neither succeeds nor fails (rare CDN stall) must not wedge
// start() forever; past this we fall back to the sync tracker.
const WORKER_INIT_TIMEOUT_MS = 15000;

/** Off-main-thread tracker: {sendFrame(video, tsMs), latest() β†’ {obs, ts}|null,
 * failed(), close()}.
 *  Latest-wins β€” sendFrame drops the frame while the worker is busy, so there is
 *  no queue growth. Resolves null when workers/bitmaps are unsupported or the
 *  worker fails to initialize (CDN imports can fail inside workers on some
 *  setups); callers then fall back to the sync createTracker() path. */
export function createWorkerTracker() {
  if (typeof Worker === "undefined" || typeof createImageBitmap === "undefined"
      || typeof OffscreenCanvas === "undefined") {
    return Promise.resolve(null);
  }
  let worker;
  try {
    worker = new Worker(new URL("./trackworker.js?v=20260729.6", import.meta.url), { type: "module" });
  } catch {
    return Promise.resolve(null);
  }
  return new Promise((resolve) => {
    let last = null;
    let busy = false;
    let ready = false;
    let failed = false;
    let closed = false;
    let timer;
    const api = {
      sendFrame(videoEl, tsMs) {
        if (failed || busy || !videoEl.videoWidth) return; // latest-wins: drop while busy
        busy = true;
        createImageBitmap(videoEl).then(
          (bitmap) => {
            if (failed) { bitmap.close(); busy = false; return; }
            try {
              worker.postMessage({ type: "frame", bitmap, ts: tsMs }, [bitmap]);
            } catch {
              bitmap.close();
              fail();
            }
          },
          () => { busy = false; },
        );
      },
      latest() {
        return last;
      },
      failed() {
        return failed;
      },
      close() {
        closed = true;
        clearTimeout(timer);
        worker.terminate();
      },
    };
    const fail = () => {
      if (closed || failed) return;
      failed = true;
      busy = false;
      clearTimeout(timer);
      worker.terminate();
      if (!ready) resolve(null);
    };
    timer = setTimeout(fail, WORKER_INIT_TIMEOUT_MS);
    // Keep this installed after ready too: an inference exception otherwise
    // leaves the latest-wins gate permanently busy with no way to recover.
    worker.onerror = fail;
    worker.onmessage = (ev) => {
      const msg = ev.data;
      if (msg.type === "result") {
        busy = false;
        last = { obs: msg.obs, ts: msg.ts };
      } else if (msg.type === "ready") {
        clearTimeout(timer);
        ready = true;
        resolve(api);
      } else if (msg.type === "fail" || msg.type === "error") {
        fail();
      }
    };
  });
}

// Face-mesh corner indices (MediaPipe topology): two eyes and the mouth ring.
const EYE_A = [33, 133];   // one eye's outer/inner corners
const EYE_B = [263, 362];  // the other eye's corners
const MOUTH = [61, 291, 13, 14, 0, 17];

/** Try to auto-locate eyes/mouth on a DRAWING (IMAGE mode). Returns
 *  {eyes:{L:[x,y],R:[x,y]}, eyeHalf, mouthBox} in canvas px, or null β€”
 *  photo-trained models often fail on hand drawings, so callers must
 *  fall back to manual clicks. */
export async function detectOnImage(canvas) {
  let landmarker = null;
  try {
    const { FaceLandmarker, FilesetResolver } = await loadTasksVision();
    const vision = await FilesetResolver.forVisionTasks(WASM_BASE);
    landmarker = await FaceLandmarker.createFromOptions(vision, {
      baseOptions: { modelAssetPath: MODEL_URL, delegate: "CPU" },
      runningMode: "IMAGE",
      numFaces: 1,
    });
    const result = landmarker.detect(canvas);
    if (!result.faceLandmarks || result.faceLandmarks.length === 0) return null;
    const lm = result.faceLandmarks[0];
    const px = (i) => [lm[i].x * canvas.width, lm[i].y * canvas.height];
    const center = (idx) => idx.map(px).reduce(([ax, ay], [x, y]) => [ax + x / idx.length, ay + y / idx.length], [0, 0]);

    const a = center(EYE_A), b = center(EYE_B);
    // Sprite 'L' is VIEWER-left on the canvas β€” assign by x, no subject-side logic needed.
    const [L, R] = a[0] <= b[0] ? [a, b] : [b, a];
    const [ax0, ax1] = EYE_A.map(px);
    const eyeHalf = Math.max(8, Math.round(Math.hypot(ax1[0] - ax0[0], ax1[1] - ax0[1]) * 0.7));

    const pts = MOUTH.map(px);
    const xs = pts.map((p) => p[0]), ys = pts.map((p) => p[1]);
    const margin = 6;
    const mouthBox = [Math.min(...xs) - margin, Math.min(...ys) - margin,
                      Math.max(...xs) + margin, Math.max(...ys) + margin].map(Math.round);
    return {
      eyes: { L: L.map(Math.round), R: R.map(Math.round) },
      eyeHalf,
      mouthBox,
      // Raw normalized landmarks β€” stored in the character manifest so the warp
      // rig can use real face geometry instead of box-synthesized rings.
      landmarks: lm.map((p) => [Number(p.x.toFixed(5)), Number(p.y.toFixed(5))]),
    };
  } catch {
    return null; // CDN/offline or detection failure β€” manual clicks still work
  } finally {
    landmarker?.close();
  }
}