@vincentt-xr/sdk 2.0.0-alpha.7 → 2.0.0-alpha.8

package/dist/types/compositor.d.ts

@@ -6,6 +6,27 @@ export type XRLayerConfig = {
     zIndex: number;
 };
 export type XRMediaSourceKind = "image" | "video";
+/** Which camera a webcam bind asks for: the front (`user`) or the rear (`environment`). */
+export type XRCameraFacing = "user" | "environment";
+/**
+ * Every kind of media a session can bind. `webcam` is a mirrored live camera
+ * (the front camera by default); `rear-camera` is a live camera drawn
+ * straight, like `image` / `video`. The mirror decision follows the kind
+ * (`isMirroredSourceKind`), so consumers never need to know the facing.
+ */
+export type XRSessionSourceKind = "webcam" | "rear-camera" | XRMediaSourceKind;
+/** Device orientation as the session sees it (from `screen.orientation`, falling back to the viewport aspect). */
+export type XROrientation = {
+    type: "portrait" | "landscape";
+    /** Clockwise screen rotation in degrees (0, 90, 180, 270); 0 when unknown. */
+    angle: 0 | 90 | 180 | 270;
+};
+/** How a webcam bind chooses and presents the camera. */
+export type XRCameraRequest = {
+    facingMode?: XRCameraFacing;
+    deviceId?: string;
+    mirrored?: boolean;
+};
 export type XRMediaLayout = {
     source: {
         width: number;

package/dist/types/holistic.d.ts

@@ -0,0 +1,121 @@
+import { NormalizedLandmark } from '@mediapipe/tasks-vision';
+import { TrackerPoint } from './common';
+/** Which hands the holistic tracker publishes. Holistic is single-person: at most one left and one right hand. */
+export type HolisticHandsOption = "both" | "left" | "right" | "none";
+export type HolisticHandSide = "left" | "right";
+export type HolisticTrackingStatus = "loading" | "ready" | "error";
+/**
+ * Inference backend. `"auto"` (default) starts on the GPU and falls back to
+ * the CPU the first time the graph fails to run there: MediaPipe's holistic
+ * face-blendshape stage has no WebGL implementation, so with
+ * `outputFaceBlendshapes` the GPU path fails on the first frame.
+ */
+export type HolisticDelegate = "auto" | "GPU" | "CPU";
+/** The backend actually running inference. */
+export type HolisticActiveDelegate = "GPU" | "CPU";
+/**
+ * Options the first acquirer passes to `XRModel.HOLISTIC_TRACKER`. The
+ * model-creation options (`modelUrl`, confidences, `outputFaceBlendshapes`)
+ * are fixed for the model's lifetime; the runtime subset
+ * (`HolisticRuntimeOptions`) can be changed live through `node.setOptions`.
+ */
+export type HolisticModelOptions = HolisticRuntimeOptions & {
+    /**
+     * Override the `.task` asset. Pre-release default: MediaPipe's published
+     * holistic landmarker on `storage.googleapis.com`, integrity-verified; the
+     * SDK CDN copy is not published yet and will become the default when it is.
+     * Set this to a self-hosted copy to drop the third-party dependency; a custom
+     * URL is fetched best-effort.
+     */
+    modelUrl?: string;
+    /** Face detector confidence, 0..1. MediaPipe default 0.5. */
+    minFaceDetectionConfidence?: number;
+    /** Face presence confidence for the landmark stage, 0..1. MediaPipe default 0.5. */
+    minFacePresenceConfidence?: number;
+    /** Pose detector confidence, 0..1. MediaPipe default 0.5. */
+    minPoseDetectionConfidence?: number;
+    /** Pose presence confidence for the landmark stage, 0..1. MediaPipe default 0.5. */
+    minPosePresenceConfidence?: number;
+    /** Hand landmark confidence, 0..1. MediaPipe default 0.5. */
+    minHandLandmarksConfidence?: number;
+    /** Emit the 52 face blendshape scores. Default `true`. */
+    outputFaceBlendshapes?: boolean;
+    /** Inference backend policy. Default `"auto"`. */
+    delegate?: HolisticDelegate;
+    /** Inference cap (fps). Read by the session at run time; overrides `perf.targetFps`. */
+    targetFps?: number;
+};
+/** Options that can change while the model is running (last writer wins). */
+export type HolisticRuntimeOptions = {
+    /** Hands to publish. Default `"both"`. */
+    hands?: HolisticHandsOption;
+    /**
+     * Longest side (px) of the frame handed to MediaPipe. The feed is
+     * downscaled to this before inference; outputs are normalized so consumers
+     * never see it. Default 768.
+     */
+    analysisMaxDimension?: number;
+    /** Keep a lost face / hand / pose for this long (ms) before dropping it. Default 250. */
+    holdMs?: number;
+    /** While held, extrapolate along the last velocity for up to this long (ms). Default 0. */
+    predictMs?: number;
+};
+/** Live face stream: landmarks in MediaPipe image space, like `FaceTracker`'s `detectionResults`. */
+export type HolisticFaceResult = {
+    /** 478 landmarks, normalized 0..1, top-left origin, Y down; `z` is tracker-relative depth. */
+    landmarks: NormalizedLandmark[];
+    /** Blendshape name → score (0..1). Empty when `outputFaceBlendshapes` is false. */
+    blendshapes: Record<string, number>;
+    /** 1 while detected; decays linearly to 0 over `holdMs` while held. */
+    confidence: number;
+    /** True when the stream is being replayed from memory after a dropout. */
+    held: boolean;
+};
+/** Live hand stream in SDK tracker coordinates, like `HandTracker`'s `coordinates`. */
+export type HolisticHandResult = {
+    /**
+     * The person's anatomical side on every source. Holistic labels follow the
+     * image, so the SDK swaps them on the mirrored webcam feed; this matches
+     * what `HandTracker` reports on the same feed.
+     */
+    handedness: HolisticHandSide;
+    /** 21 joints, center origin, X right, Y up, Z toward the viewer. */
+    landmarks: TrackerPoint[];
+    /** 21 joints in MediaPipe's metric world space (metres, wrist-relative, Y down). Not converted. */
+    worldLandmarks: TrackerPoint[];
+    /** 1 while detected; decays linearly to 0 over `holdMs` while held. */
+    confidence: number;
+    held: boolean;
+};
+/**
+ * Live pose stream in SDK tracker coordinates, like `BodyTracker`'s
+ * `getLandmarks2D()`. Unlike `BodyTracker`, left/right indices are the
+ * person's anatomical sides on every source (swapped on the mirrored webcam
+ * feed), so they agree with `leftHand` / `rightHand`.
+ */
+export type HolisticPoseResult = {
+    /** 33 landmarks, center origin, X right, Y up, Z toward the viewer. */
+    landmarks: TrackerPoint[];
+    /** 33 landmarks in MediaPipe's metric world space (metres, hip-relative, Y down). Not converted. */
+    worldLandmarks: TrackerPoint[];
+    /** Per-landmark visibility from MediaPipe (0..1), aligned with `landmarks`. */
+    visibility: number[];
+    /** Mean landmark visibility while detected; decays linearly to 0 over `holdMs` while held. */
+    confidence: number;
+    held: boolean;
+};
+/**
+ * The unified holistic result. One object per model instance, mutated in
+ * place every frame; read it inside `useFrame` and treat it as read-only.
+ * A stream is `null` when it is neither detected nor held.
+ */
+export type HolisticResult = {
+    /** Bumped on every publish; `0` until the first inference. */
+    version: number;
+    /** `performance.now()` of the last publish. */
+    timestampMs: number;
+    face: HolisticFaceResult | null;
+    leftHand: HolisticHandResult | null;
+    rightHand: HolisticHandResult | null;
+    pose: HolisticPoseResult | null;
+};

package/dist/types/index.d.ts

@@ -3,6 +3,7 @@ export * from './session';
 export * from './body';
 export * from './face';
 export * from './headBinding';
+export * from './holistic';
 export * from './model';
 export * from './segmentation';
 export * from './common';

package/dist/types/model.d.ts

@@ -1,4 +1,4 @@
-import { XRMediaSourceKind } from './compositor';
+import { XRSessionSourceKind } from './compositor';
 /**
  * Session-level performance config, set via `<XRProvider perf={...}>`. Trades
  * tracking fidelity for lower GPU/CPU load and heat. Every field is optional;
@@ -29,12 +29,30 @@ export type XRPerfConfig = {
      * the tight native-selfie framing users expect; set `1` for the widest FOV.
      */
     cameraZoom?: number;
+    /**
+     * Multiplier (0.25..1) on every tracker's analysis resolution. Lower =
+     * cheaper inference, coarser landmarks. The adaptive-quality controller
+     * drives it per tier; set it directly only with `quality={false}`.
+     */
+    analysisScale?: number;
+};
+/**
+ * Session lifecycle behaviour, set via `<XRProvider lifecycle={...}>`.
+ */
+export type XRLifecycleConfig = {
+    /**
+     * Pause the whole session (blit loop, every model, the video element) while
+     * the tab is hidden and resume it when it returns. Off by default: a hidden
+     * tab already throttles RAF, and some flows (recording, picture-in-picture)
+     * must keep running. Turn it on for battery-sensitive single-page apps.
+     */
+    pauseWhenHidden?: boolean;
 };
 export type RunModelParams = {
     sourceCanvas: OffscreenCanvas | HTMLCanvasElement;
     renderWidth: number;
     renderHeight: number;
-    getSource?: () => "webcam" | XRMediaSourceKind | undefined;
+    getSource?: () => XRSessionSourceKind | undefined;
     /**
      * Cap the model's inference rate to this many frames per second. The model's
      * RAF loop still runs at display rate, but the (expensive) MediaPipe detect
@@ -43,4 +61,6 @@ export type RunModelParams = {
      * smoothness. See `session.perf.targetFps` and the per-tracker `targetFps`.
      */
     targetFps?: number;
+    /** Multiplier on the model's analysis resolution (0.25..1); default 1. */
+    analysisScale?: number;
 };

package/dist/types/nodes.d.ts

@@ -1,10 +1,12 @@
-import { FaceLandmarkerResult } from '@mediapipe/tasks-vision';
+import { FaceLandmarkerResult, HolisticLandmarkerResult } from '@mediapipe/tasks-vision';
 import { Texture, Camera } from 'three';
 import { TrackerPoint } from './common';
 import { Face, FaceRawTrackingData } from './face';
 import { GestureData } from './gesture-tracking';
 import { HandData } from './hand-tracking';
-import { SegmentationData } from './segmentation';
+import { HolisticActiveDelegate, HolisticResult, HolisticRuntimeOptions, HolisticTrackingStatus } from './holistic';
+import { SemanticMask } from '../segmentation2d/types';
+import { SegmentationData, SemanticSegmentationStatus } from './segmentation';
 export type XRModelNodeSegmentation = {
     backgroundSegmentationMask?: Texture;
     getMaskCanvas: () => HTMLCanvasElement | null;
@@ -16,6 +18,29 @@ export type XRModelNodeSegmentation = {
      */
     setInputSource: (source: CanvasImageSource | null | undefined) => void;
 };
+/** Live state of `XRModel.SEMANTIC_SEGMENTATION`; functions read the model's current values. */
+export type XRModelNodeSemanticSegmentation = {
+    /** Stable mask object in feed space, mutated in place; `version` 0 until the first inference. */
+    getMask: () => SemanticMask;
+    getStatus: () => SemanticSegmentationStatus;
+    getError: () => Error | null;
+};
+/**
+ * Live state of `XRModel.HOLISTIC_TRACKER`. One node object per model
+ * instance; every function reads the current value.
+ */
+export type XRModelNodeHolisticTracking = {
+    /** Unified, dropout-bridged result. Stable object mutated in place; `version` 0 until the first inference. */
+    getResult: () => HolisticResult;
+    /** The last raw MediaPipe result, unbridged and unconverted. */
+    getRawResult: () => HolisticLandmarkerResult | undefined;
+    getStatus: () => HolisticTrackingStatus;
+    getError: () => Error | null;
+    /** The backend running inference; `"CPU"` after an automatic fallback. */
+    getDelegate: () => HolisticActiveDelegate;
+    /** Change the runtime-tunable options without re-initializing the model. */
+    setOptions: (options: HolisticRuntimeOptions) => void;
+};
 export type XRModelNodeBodyTracking = {
     getLandmarks2D: (bodyIndex?: number) => TrackerPoint[] | undefined;
 };

package/dist/types/segmentation.d.ts

@@ -29,3 +29,22 @@ export type SegmentationModelOptions = {
      */
     quality?: SegmentationQuality;
 };
+/**
+ * Category-segmentation presets served by `XRModel.SEMANTIC_SEGMENTATION`.
+ * `hair` runs MediaPipe's Hair Segmenter (`bg` / `hair`); `multiclass` runs
+ * the Selfie Multiclass model (`bg`, `hair`, `body_skin`, `face_skin`,
+ * `clothes`, `others`). One preset per session: the first acquirer decides.
+ */
+export declare const SEMANTIC_SEGMENTATION_PRESETS: readonly ["hair", "multiclass"];
+export type SemanticSegmentationPreset = (typeof SEMANTIC_SEGMENTATION_PRESETS)[number];
+export type SemanticSegmentationModelOptions = {
+    preset?: SemanticSegmentationPreset;
+    /**
+     * Override the model asset URL. Defaults to the SDK CDN path for the
+     * preset, which is integrity-verified; a custom URL is fetched best-effort.
+     */
+    modelUrl?: string;
+    /** Process-canvas scale relative to the feed: low 0.5, medium 0.75, high 1. Default medium. */
+    quality?: SegmentationQuality;
+};
+export type SemanticSegmentationStatus = "loading" | "ready" | "error";

package/dist/types/session.d.ts

@@ -1,8 +1,9 @@
 import { XRMediaSource, XRModel } from '../constants';
-import { XRMediaSourceKind } from './compositor';
+import { XRCameraFacing, XRCameraRequest, XRMediaSourceKind } from './compositor';
 import { BodyModelOptions } from './body';
 import { FaceModelOptions } from './face';
-import { SegmentationModelOptions } from './segmentation';
+import { HolisticModelOptions } from './holistic';
+import { SegmentationModelOptions, SemanticSegmentationModelOptions } from './segmentation';
 export declare enum SessionStatus {
     ACTIVE = "ACTIVE",
     DISABLED = "DISABLED"
@@ -18,6 +19,19 @@ export type Session = {
 type SetMediaSourceWebcamParams = {
     source: XRMediaSource.WEBCAM;
     sourceKind?: "webcam";
+    /**
+     * Which camera to open: `user` (front, default) or `environment` (rear).
+     * Binding again with a different facing swaps the camera mid-session.
+     */
+    facingMode?: XRCameraFacing;
+    /** A specific camera from `enumerateDevices()`; beats `facingMode` when set. */
+    deviceId?: string;
+    /**
+     * Mirror override. Default: the front camera is mirrored (`sourceKind`
+     * `"webcam"`), the rear camera is drawn straight (`sourceKind`
+     * `"rear-camera"`). Set explicitly for a camera whose facing is unknown.
+     */
+    mirrored?: boolean;
 };
 type SetMediaSourceStreamParams = {
     source: XRMediaSource.STREAM;
@@ -44,9 +58,15 @@ export type RegisterXRPipelineParams = Array<{
 } | {
     model: XRModel.FACE_TRACKER;
     options?: FaceModelOptions;
+} | {
+    model: XRModel.HOLISTIC_TRACKER;
+    options?: HolisticModelOptions;
 } | {
     model: XRModel.BACKGROUND_SEGMENTATION;
     options?: SegmentationModelOptions;
+} | {
+    model: XRModel.SEMANTIC_SEGMENTATION;
+    options?: SemanticSegmentationModelOptions;
 }>;
 /**
  * Public XR session surface. Returned by `useXRContext().session` and used
@@ -62,7 +82,10 @@ export type RegisterXRPipelineParams = Array<{
  * Lifecycle:
  *
  *     idle ──setMediaSource──> media-bound ──start──> running ──stop──> stopped
- *                                   ↑                                       │
+ *                                   ↑                 │    ↑                │
+ *                                   │           pause │    │ resume          │
+ *                                   │                 ▼    │                │
+ *                                   │               paused ─┘               │
  *                                   └────────────reset──────────────────────┘
  */
 export type XRSession = {
@@ -83,6 +106,21 @@ export type XRSession = {
      * `stop()` + clear the model registry. Returns the session to `idle`.
      */
     reset: () => Promise<void>;
+    /**
+     * Suspend a `running` session without releasing the camera: the feed loop
+     * stops, every model pauses, the video element pauses. No-op in any other
+     * state. `resume()` (or `start()`) continues with no rebind.
+     */
+    pause: () => void;
+    /** Continue a `paused` session. No-op in any other state. */
+    resume: () => Promise<void>;
+    /**
+     * Open another camera (`facingMode` / `deviceId`, optional `mirrored`
+     * override) and keep going: models, lifecycle and scene stay; a
+     * `media-bound` / `stopped` session is started. Resolves to the facing the
+     * session ended up with (a single-camera device answers with what it has).
+     */
+    switchCamera: (request?: XRCameraRequest) => Promise<XRCameraFacing>;
     /**
      * @deprecated Prefer `useXRModel(modelName, options)`. Kept as an
      * additive shim for backwards compatibility — multiple calls are

package/dist/useFaceInfo-Blf7FCu3.js

@@ -0,0 +1,303 @@
+import { useMemo as I, useEffect as N, useCallback as k, useRef as U } from "react";
+import { Vector2 as X, Raycaster as S, Ray as Y, Vector3 as L } from "three";
+import { D as d, a as y } from "./screenTransform-BenfCtZt.js";
+import { n as j, t as A, X as p } from "./index-CnGox6Xr.js";
+import { useThree as D } from "@react-three/fiber";
+import { g as B, u as J } from "./getFaceInfo-BkZfaQs-.js";
+import { u as V } from "./index-CsUdQ-pn.js";
+const a = (e, n) => {
+  if (!Number.isFinite(e) || e <= 0) throw new RangeError(`${n} must be positive`);
+  return e;
+}, q = (e, n) => {
+  const t = a(n.width, "root.width"), s = a(n.height, "root.height");
+  return !Number.isFinite(e.width) || !Number.isFinite(e.height) || e.width <= 0 || e.height <= 0 ? 0 : Math.min(e.width / t, e.height / s);
+}, b = (e, n, t, s) => n.width <= 0 || n.height <= 0 ? Object.assign(s, { x: 0, y: 0 }) : (a(t.width, "root.width"), a(t.height, "root.height"), s.x = (e.x / n.width - 0.5) * t.width, s.y = (0.5 - e.y / n.height) * t.height, s), P = (e, n, t, s) => n.width <= 0 || n.height <= 0 ? Object.assign(s, { x: 0, y: 0 }) : (a(t.width, "root.width"), a(t.height, "root.height"), s.x = (e.x / t.width + 0.5) * n.width, s.y = (0.5 - e.y / t.height) * n.height, s), O = ({
+  parentSize: e,
+  referencePixelsPerUnit: n = y,
+  root: t = d
+}) => q(e, t) * (y / a(n, "referencePixelsPerUnit")), re = (e, n) => e * O(n), ce = (e, n) => {
+  const t = O(n);
+  return t === 0 ? 0 : e / t;
+}, oe = ({
+  point: e,
+  parentSize: n,
+  referencePixelsPerUnit: t = y,
+  root: s = d,
+  target: i = { x: 0, y: 0 }
+}) => {
+  if (n.width <= 0 || n.height <= 0) return Object.assign(i, { x: 0, y: 0 });
+  const r = y / a(t, "referencePixelsPerUnit");
+  return i.x = e.x * (n.width / a(s.width, "root.width")) * r, i.y = e.y * (n.height / a(s.height, "root.height")) * r, i;
+}, he = ({
+  point: e,
+  parentSize: n,
+  referencePixelsPerUnit: t = y,
+  root: s = d,
+  target: i = { x: 0, y: 0 }
+}) => {
+  if (n.width <= 0 || n.height <= 0) return Object.assign(i, { x: 0, y: 0 });
+  const r = y / a(t, "referencePixelsPerUnit");
+  return i.x = e.x / (n.width / a(s.width, "root.width") * r), i.y = e.y / (n.height / a(s.height, "root.height") * r), i;
+}, xe = ({
+  point: e,
+  rect: n,
+  root: t = d,
+  target: s = { x: 0, y: 0 }
+}) => b({ x: e.x - n.left, y: e.y - n.top }, n, t, s), ae = ({
+  point: e,
+  rect: n,
+  root: t = d,
+  target: s = { x: 0, y: 0 }
+}) => (P(e, n, t, s), s.x += n.left, s.y += n.top, s), _ = ({
+  point: e,
+  root: n = d,
+  target: t = { x: 0, y: 0 },
+  viewportSize: s
+}) => b(
+  { x: e.x * s.width, y: e.y * s.height },
+  s,
+  n,
+  t
+), K = ({
+  point: e,
+  root: n = d,
+  target: t = { x: 0, y: 0 },
+  viewportSize: s
+}) => s.width <= 0 || s.height <= 0 ? Object.assign(t, { x: 0, y: 0 }) : (P(e, s, n, t), t.x /= s.width, t.y /= s.height, t), de = ({
+  point: e,
+  root: n = d,
+  target: t = { x: 0, y: 0 },
+  viewportSize: s
+}) => _({
+  point: { x: (e.x + 1) / 2, y: (1 - e.y) / 2 },
+  viewportSize: s,
+  root: n,
+  target: t
+}), le = ({
+  layout: e,
+  mirrored: n = !1,
+  point: t,
+  root: s = d,
+  target: i = { x: 0, y: 0 }
+}) => {
+  if (e.scale <= 0) return Object.assign(i, { x: 0, y: 0 });
+  const r = e.offsetX + t.x * e.scale;
+  return b(
+    {
+      x: n ? e.destination.width - r : r,
+      y: e.offsetY + t.y * e.scale
+    },
+    e.destination,
+    s,
+    i
+  );
+}, ue = ({
+  layout: e,
+  mirrored: n = !1,
+  point: t,
+  root: s = d,
+  target: i = { x: 0, y: 0 }
+}) => {
+  if (e.scale <= 0) return Object.assign(i, { x: 0, y: 0 });
+  P(t, e.destination, s, i);
+  const r = n ? e.destination.width - i.x : i.x;
+  return i.x = (r - e.offsetX) / e.scale, i.y = (i.y - e.offsetY) / e.scale, i;
+}, w = new L(), g = new X(), R = new S(), fe = ({
+  camera: e,
+  point: n,
+  root: t = d,
+  target: s = { x: 0, y: 0, depth: 0 },
+  viewportSize: i
+}) => (a(t.width, "root.width"), a(t.height, "root.height"), w.set(n.x, n.y, n.z).project(e), s.x = w.x * t.width / 2, s.y = w.y * t.height / 2, s.depth = w.z, s), ye = ({
+  camera: e,
+  point: n,
+  root: t = d,
+  target: s = new Y(),
+  viewportSize: i
+}) => (a(t.width, "root.width"), a(t.height, "root.height"), g.set(n.x * 2 / t.width, n.y * 2 / t.height), R.setFromCamera(g, e), s.copy(R.ray)), me = (e, n) => j({ x: e.x, y: e.y, z: e.z ?? 0 }, n), H = (e, n = { x: 0, y: 0 }) => {
+  const t = A(e);
+  return n.x = t.x, n.y = t.y, n.z = t.z, n;
+}, we = (e, n, t = { x: 0, y: 0 }) => (t.x = e.x * n.width, t.y = e.y * n.height, t), Te = (e, n, t = { x: 0, y: 0 }) => n.width <= 0 || n.height <= 0 ? Object.assign(t, { x: 0, y: 0 }) : (t.x = e.x / n.width, t.y = e.y / n.height, t), be = (e, n, t = {}) => _({
+  point: e,
+  viewportSize: n,
+  root: t.root,
+  target: t.target
+}), Pe = (e, n, t = {}) => {
+  const s = t.target ?? { x: 0, y: 0 };
+  return K({ point: e, viewportSize: n, root: t.root, target: s }), s;
+}, W = { selection: "index", index: 0 }, $ = Object.freeze({
+  face: { max: 10, default: 1, option: "numFaces" },
+  hand: { max: 2, default: 2, option: null },
+  gesture: { max: 1, default: 1, option: null },
+  body: { max: 1, default: 1, option: null },
+  holistic: { max: 1, default: 1, option: null }
+}), C = { center: null }, f = (e, n) => Math.hypot(e.x - n.x, e.y - n.y), M = (e, n, t) => {
+  if (!e.length) return null;
+  switch (n) {
+    case "largest":
+      return e.reduce((s, i) => i.size > s.size ? i : s);
+    case "nearest-center":
+      return e.reduce(
+        (s, i) => f(i.center, { x: 0.5, y: 0.5 }) < f(s.center, { x: 0.5, y: 0.5 }) ? i : s
+      );
+    case "leftmost":
+      return e.reduce((s, i) => i.center.x < s.center.x ? i : s);
+    case "rightmost":
+      return e.reduce((s, i) => i.center.x > s.center.x ? i : s);
+    default:
+      return e.find((s) => s.index === t) ?? null;
+  }
+}, G = (e, n = W, t = C) => {
+  const s = n.index ?? 0;
+  if (n.selection !== "sticky") {
+    const c = M(e, n.selection, s);
+    return {
+      index: (c == null ? void 0 : c.index) ?? null,
+      reason: c ? n.selection : "none",
+      memory: c ? { center: { ...c.center }, label: c.label } : t
+    };
+  }
+  const i = n.stickyRadius ?? 0.2;
+  if (t.center) {
+    const c = t.center, x = e.filter((l) => f(l.center, c) <= i).sort((l, m) => f(l.center, c) - f(m.center, c)), h = (t.label ? x.find((l) => l.label === t.label) : void 0) ?? x[0];
+    if (h)
+      return {
+        index: h.index,
+        reason: "sticky",
+        memory: { center: { ...h.center }, label: h.label }
+      };
+  }
+  const r = M(e, n.fallback ?? "largest", s);
+  return {
+    index: (r == null ? void 0 : r.index) ?? null,
+    reason: r ? "sticky-fallback" : "none",
+    // Keep the old memory on a miss so a brief dropout does not forget the subject.
+    memory: r ? { center: { ...r.center }, label: r.label } : t
+  };
+}, Fe = (e) => e.map((n, t) => {
+  let s = 1 / 0, i = -1 / 0, r = 1 / 0, c = -1 / 0;
+  for (const o of n)
+    o.x < s && (s = o.x), o.x > i && (i = o.x), o.y < r && (r = o.y), o.y > c && (c = o.y);
+  const x = n.length > 0 && Number.isFinite(s);
+  return {
+    index: t,
+    center: x ? { x: (s + i) / 2, y: (r + c) / 2 } : { x: 0.5, y: 0.5 },
+    size: x ? i - s : 0
+  };
+}), Ee = (e) => e.map((n, t) => {
+  const s = (n.coordinates ?? []).map((h) => H(h));
+  let i = 1 / 0, r = -1 / 0, c = 1 / 0, x = -1 / 0;
+  for (const h of s)
+    h.x < i && (i = h.x), h.x > r && (r = h.x), h.y < c && (c = h.y), h.y > x && (x = h.y);
+  const o = s.length > 0;
+  return {
+    index: t,
+    center: o ? { x: (i + r) / 2, y: (c + x) / 2 } : { x: 0.5, y: 0.5 },
+    size: o ? Math.max(r - i, x - c) : 0,
+    ...n.label ? { label: n.label } : {}
+  };
+}), ge = (e, n, t = 0.35) => {
+  const s = [];
+  for (const c of e)
+    for (const x of n) {
+      const o = f(c.center, x.center);
+      o <= t && s.push({ primaryIndex: c.index, secondaryIndex: x.index, distance: o });
+    }
+  s.sort((c, x) => c.distance - x.distance);
+  const i = /* @__PURE__ */ new Set(), r = /* @__PURE__ */ new Set();
+  return s.filter((c) => i.has(c.primaryIndex) || r.has(c.secondaryIndex) ? !1 : (i.add(c.primaryIndex), r.add(c.secondaryIndex), !0));
+}, Q = ({
+  active: e = !0,
+  holdMs: n = 250,
+  numFaces: t = 1,
+  targetFps: s,
+  throttleMs: i = 16
+} = {}) => {
+  const { camera: r } = D(), c = Number.isFinite(t) ? Math.min($.face.max, Math.max(1, Math.round(t))) : 1, x = s !== void 0 && Number.isFinite(s) && s > 0 ? s : void 0, o = I(
+    () => ({
+      numFaces: c,
+      ...x ? { targetFps: x } : {}
+    }),
+    [c, x]
+  ), { node: h, ready: l } = V(
+    p.FACE_TRACKER,
+    o,
+    e
+  );
+  N(() => {
+    if (h)
+      for (let u = 0; u < o.numFaces; u += 1)
+        h.setCamera(r, u);
+  }, [o.numFaces, r, h]);
+  const m = k(
+    (u) => {
+      if (!h) return [];
+      const F = [];
+      for (let T = 0; T < u.faceLandmarks.length; T += 1) {
+        const E = B(h, u, T);
+        E && F.push(E);
+      }
+      return F;
+    },
+    [h]
+  );
+  return J({
+    enabled: e && l,
+    holdMs: Number.isFinite(n) ? Math.max(0, n) : 250,
+    node: h,
+    prepare: m,
+    throttleMs: Number.isFinite(i) ? Math.max(0, i) : 16
+  });
+}, Z = 4, Re = ({
+  faceIndex: e = 0,
+  numFaces: n,
+  subject: t,
+  ...s
+} = {}) => {
+  const i = Number.isFinite(e) ? Math.max(0, Math.round(e)) : 0, r = Q({
+    ...s,
+    numFaces: Math.max(n ?? (t ? Z : i + 1), i + 1)
+  }), c = U(C), x = JSON.stringify(t ?? null), o = I(() => {
+    if (!t) return null;
+    const h = r.map((l, m) => ({
+      index: m,
+      center: { x: l.bounds.centerX, y: l.bounds.centerY },
+      size: l.bounds.width
+    }));
+    return G(h, t, c.current);
+  }, [r, x]);
+  return N(() => {
+    o && (c.current = o.memory);
+  }, [o]), t ? o && o.index !== null ? r[o.index] ?? null : null : r[i] ?? null;
+};
+export {
+  W as D,
+  C as E,
+  $ as X,
+  ge as a,
+  Z as b,
+  Re as c,
+  xe as d,
+  re as e,
+  Fe as f,
+  ae as g,
+  ue as h,
+  K as i,
+  Ee as j,
+  oe as k,
+  ye as l,
+  le as m,
+  _ as n,
+  he as o,
+  ce as p,
+  be as q,
+  me as r,
+  G as s,
+  de as t,
+  Q as u,
+  we as v,
+  fe as w,
+  Pe as x,
+  H as y,
+  Te as z
+};

package/dist/useFaceInfo-C5mO8NQY.cjs

@@ -0,0 +1 @@
+"use strict";const T=require("react"),R=require("three"),l=require("./screenTransform-BiXdpVf0.cjs"),F=require("./index-CeVkgsag.cjs"),j=require("@react-three/fiber"),w=require("./getFaceInfo-ahPxQFGu.cjs"),q=require("./index-OCPjOaeP.cjs"),h=(e,n)=>{if(!Number.isFinite(e)||e<=0)throw new RangeError(`${n} must be positive`);return e},B=(e,n)=>{const t=h(n.width,"root.width"),i=h(n.height,"root.height");return!Number.isFinite(e.width)||!Number.isFinite(e.height)||e.width<=0||e.height<=0?0:Math.min(e.width/t,e.height/i)},m=(e,n,t,i)=>n.width<=0||n.height<=0?Object.assign(i,{x:0,y:0}):(h(t.width,"root.width"),h(t.height,"root.height"),i.x=(e.x/n.width-.5)*t.width,i.y=(.5-e.y/n.height)*t.height,i),P=(e,n,t,i)=>n.width<=0||n.height<=0?Object.assign(i,{x:0,y:0}):(h(t.width,"root.width"),h(t.height,"root.height"),i.x=(e.x/t.width+.5)*n.width,i.y=(.5-e.y/t.height)*n.height,i),I=({parentSize:e,referencePixelsPerUnit:n=l.DEFAULT_REFERENCE_PIXELS_PER_UNIT,root:t=l.DEFAULT_SCREEN_TRANSFORM_ROOT})=>B(e,t)*(l.DEFAULT_REFERENCE_PIXELS_PER_UNIT/h(n,"referencePixelsPerUnit")),J=(e,n)=>e*I(n),V=(e,n)=>{const t=I(n);return t===0?0:e/t},p=({point:e,parentSize:n,referencePixelsPerUnit:t=l.DEFAULT_REFERENCE_PIXELS_PER_UNIT,root:i=l.DEFAULT_SCREEN_TRANSFORM_ROOT,target:s={x:0,y:0}})=>{if(n.width<=0||n.height<=0)return Object.assign(s,{x:0,y:0});const r=l.DEFAULT_REFERENCE_PIXELS_PER_UNIT/h(t,"referencePixelsPerUnit");return s.x=e.x*(n.width/h(i.width,"root.width"))*r,s.y=e.y*(n.height/h(i.height,"root.height"))*r,s},H=({point:e,parentSize:n,referencePixelsPerUnit:t=l.DEFAULT_REFERENCE_PIXELS_PER_UNIT,root:i=l.DEFAULT_SCREEN_TRANSFORM_ROOT,target:s={x:0,y:0}})=>{if(n.width<=0||n.height<=0)return Object.assign(s,{x:0,y:0});const r=l.DEFAULT_REFERENCE_PIXELS_PER_UNIT/h(t,"referencePixelsPerUnit");return s.x=e.x/(n.width/h(i.width,"root.width")*r),s.y=e.y/(n.height/h(i.height,"root.height")*r),s},K=({point:e,rect:n,root:t=l.DEFAULT_SCREEN_TRANSFORM_ROOT,target:i={x:0,y:0}})=>m({x:e.x-n.left,y:e.y-n.top},n,t,i),$=({point:e,rect:n,root:t=l.DEFAULT_SCREEN_TRANSFORM_ROOT,target:i={x:0,y:0}})=>(P(e,n,t,i),i.x+=n.left,i.y+=n.top,i),S=({point:e,root:n=l.DEFAULT_SCREEN_TRANSFORM_ROOT,target:t={x:0,y:0},viewportSize:i})=>m({x:e.x*i.width,y:e.y*i.height},i,n,t),L=({point:e,root:n=l.DEFAULT_SCREEN_TRANSFORM_ROOT,target:t={x:0,y:0},viewportSize:i})=>i.width<=0||i.height<=0?Object.assign(t,{x:0,y:0}):(P(e,i,n,t),t.x/=i.width,t.y/=i.height,t),W=({point:e,root:n=l.DEFAULT_SCREEN_TRANSFORM_ROOT,target:t={x:0,y:0},viewportSize:i})=>S({point:{x:(e.x+1)/2,y:(1-e.y)/2},viewportSize:i,root:n,target:t}),z=({layout:e,mirrored:n=!1,point:t,root:i=l.DEFAULT_SCREEN_TRANSFORM_ROOT,target:s={x:0,y:0}})=>{if(e.scale<=0)return Object.assign(s,{x:0,y:0});const r=e.offsetX+t.x*e.scale;return m({x:n?e.destination.width-r:r,y:e.offsetY+t.y*e.scale},e.destination,i,s)},G=({layout:e,mirrored:n=!1,point:t,root:i=l.DEFAULT_SCREEN_TRANSFORM_ROOT,target:s={x:0,y:0}})=>{if(e.scale<=0)return Object.assign(s,{x:0,y:0});P(t,e.destination,i,s);const r=n?e.destination.width-s.x:s.x;return s.x=(r-e.offsetX)/e.scale,s.y=(s.y-e.offsetY)/e.scale,s},y=new R.Vector3,C=new R.Vector2,A=new R.Raycaster,Q=({camera:e,point:n,root:t=l.DEFAULT_SCREEN_TRANSFORM_ROOT,target:i={x:0,y:0,depth:0},viewportSize:s})=>(h(t.width,"root.width"),h(t.height,"root.height"),y.set(n.x,n.y,n.z).project(e),i.x=y.x*t.width/2,i.y=y.y*t.height/2,i.depth=y.z,i),Z=({camera:e,point:n,root:t=l.DEFAULT_SCREEN_TRANSFORM_ROOT,target:i=new R.Ray,viewportSize:s})=>(h(t.width,"root.width"),h(t.height,"root.height"),C.set(n.x*2/t.width,n.y*2/t.height),A.setFromCamera(C,e),i.copy(A.ray)),v=(e,n)=>F.normalizedTrackerPointToTrackerPoint({x:e.x,y:e.y,z:e.z??0},n),b=(e,n={x:0,y:0})=>{const t=F.trackerPointToNormalizedTrackerPoint(e);return n.x=t.x,n.y=t.y,n.z=t.z,n},ee=(e,n,t={x:0,y:0})=>(t.x=e.x*n.width,t.y=e.y*n.height,t),ne=(e,n,t={x:0,y:0})=>n.width<=0||n.height<=0?Object.assign(t,{x:0,y:0}):(t.x=e.x/n.width,t.y=e.y/n.height,t),te=(e,n,t={})=>S({point:e,viewportSize:n,root:t.root,target:t.target}),ie=(e,n,t={})=>{const i=t.target??{x:0,y:0};return L({point:e,viewportSize:n,root:t.root,target:i}),i},D={selection:"index",index:0},g=Object.freeze({face:{max:10,default:1,option:"numFaces"},hand:{max:2,default:2,option:null},gesture:{max:1,default:1,option:null},body:{max:1,default:1,option:null},holistic:{max:1,default:1,option:null}}),O={center:null},f=(e,n)=>Math.hypot(e.x-n.x,e.y-n.y),M=(e,n,t)=>{if(!e.length)return null;switch(n){case"largest":return e.reduce((i,s)=>s.size>i.size?s:i);case"nearest-center":return e.reduce((i,s)=>f(s.center,{x:.5,y:.5})<f(i.center,{x:.5,y:.5})?s:i);case"leftmost":return e.reduce((i,s)=>s.center.x<i.center.x?s:i);case"rightmost":return e.reduce((i,s)=>s.center.x>i.center.x?s:i);default:return e.find(i=>i.index===t)??null}},X=(e,n=D,t=O)=>{const i=n.index??0;if(n.selection!=="sticky"){const c=M(e,n.selection,i);return{index:(c==null?void 0:c.index)??null,reason:c?n.selection:"none",memory:c?{center:{...c.center},label:c.label}:t}}const s=n.stickyRadius??.2;if(t.center){const c=t.center,d=e.filter(a=>f(a.center,c)<=s).sort((a,E)=>f(a.center,c)-f(E.center,c)),x=(t.label?d.find(a=>a.label===t.label):void 0)??d[0];if(x)return{index:x.index,reason:"sticky",memory:{center:{...x.center},label:x.label}}}const r=M(e,n.fallback??"largest",i);return{index:(r==null?void 0:r.index)??null,reason:r?"sticky-fallback":"none",memory:r?{center:{...r.center},label:r.label}:t}},se=e=>e.map((n,t)=>{let i=1/0,s=-1/0,r=1/0,c=-1/0;for(const o of n)o.x<i&&(i=o.x),o.x>s&&(s=o.x),o.y<r&&(r=o.y),o.y>c&&(c=o.y);const d=n.length>0&&Number.isFinite(i);return{index:t,center:d?{x:(i+s)/2,y:(r+c)/2}:{x:.5,y:.5},size:d?s-i:0}}),re=e=>e.map((n,t)=>{const i=(n.coordinates??[]).map(x=>b(x));let s=1/0,r=-1/0,c=1/0,d=-1/0;for(const x of i)x.x<s&&(s=x.x),x.x>r&&(r=x.x),x.y<c&&(c=x.y),x.y>d&&(d=x.y);const o=i.length>0;return{index:t,center:o?{x:(s+r)/2,y:(c+d)/2}:{x:.5,y:.5},size:o?Math.max(r-s,d-c):0,...n.label?{label:n.label}:{}}}),ce=(e,n,t=.35)=>{const i=[];for(const c of e)for(const d of n){const o=f(c.center,d.center);o<=t&&i.push({primaryIndex:c.index,secondaryIndex:d.index,distance:o})}i.sort((c,d)=>c.distance-d.distance);const s=new Set,r=new Set;return i.filter(c=>s.has(c.primaryIndex)||r.has(c.secondaryIndex)?!1:(s.add(c.primaryIndex),r.add(c.secondaryIndex),!0))},k=({active:e=!0,holdMs:n=250,numFaces:t=1,targetFps:i,throttleMs:s=16}={})=>{const{camera:r}=j.useThree(),c=Number.isFinite(t)?Math.min(g.face.max,Math.max(1,Math.round(t))):1,d=i!==void 0&&Number.isFinite(i)&&i>0?i:void 0,o=T.useMemo(()=>({numFaces:c,...d?{targetFps:d}:{}}),[c,d]),{node:x,ready:a}=q.useXRModel(F.XRModel.FACE_TRACKER,o,e);T.useEffect(()=>{if(x)for(let u=0;u<o.numFaces;u+=1)x.setCamera(r,u)},[o.numFaces,r,x]);const E=T.useCallback(u=>{if(!x)return[];const N=[];for(let _=0;_<u.faceLandmarks.length;_+=1){const U=w.getFaceInfoFromResults(x,u,_);U&&N.push(U)}return N},[x]);return w.useFaceResultReader({enabled:e&&a,holdMs:Number.isFinite(n)?Math.max(0,n):250,node:x,prepare:E,throttleMs:Number.isFinite(s)?Math.max(0,s):16})},Y=4,oe=({faceIndex:e=0,numFaces:n,subject:t,...i}={})=>{const s=Number.isFinite(e)?Math.max(0,Math.round(e)):0,r=k({...i,numFaces:Math.max(n??(t?Y:s+1),s+1)}),c=T.useRef(O),d=JSON.stringify(t??null),o=T.useMemo(()=>{if(!t)return null;const x=r.map((a,E)=>({index:E,center:{x:a.bounds.centerX,y:a.bounds.centerY},size:a.bounds.width}));return X(x,t,c.current)},[r,d]);return T.useEffect(()=>{o&&(c.current=o.memory)},[o]),t?o&&o.index!==null?r[o.index]??null:null:r[s]??null};exports.DEFAULT_POLICY_NUM_FACES=Y;exports.DEFAULT_SUBJECT_POLICY=D;exports.EMPTY_SUBJECT_MEMORY=O;exports.XR_MAX_SUBJECTS=g;exports.associateSubjects=ce;exports.clientToScreenPixels=K;exports.faceCandidates=se;exports.feedToScreenPixels=te;exports.feedToTracker=v;exports.feedToViewport=ee;exports.jointCandidates=re;exports.mediaToScreenPixels=z;exports.normalizedToScreenPixels=S;exports.screenPixelsToClient=$;exports.screenPixelsToFeed=ie;exports.screenPixelsToMedia=G;exports.screenPixelsToNormalized=L;exports.screenPixelsToScreenSpace=p;exports.screenPixelsToUnits=J;exports.screenPointToRay=Z;exports.screenSpaceToScreenPixels=H;exports.screenUnitsToPixels=V;exports.selectSubject=X;exports.trackerPointToScreenPixels=W;exports.trackerToFeed=b;exports.useFaceInfo=oe;exports.useFaceInfos=k;exports.viewportToFeed=ne;exports.worldToScreenPixels=Q;