fix: FOV applies to the image's long axis (iOS gives portrait streams — focal was 1.78x short); cap detection canvas long side (portrait was 1.6MP not 0.5MP); show duplicate detections in HUD

public/js/exhibit.js

@@ -59,7 +59,7 @@ let detector, poseEst, fuser, cvSolver;
 let gradients = {}, modelManifest = null, characters = { defaults: {}, byId: {} };
 let ghostHeight = 4; // cm, from scene.ghostHeightCm
 const activeGhosts = new Map(); // uid -> { rec, group }
-let tracking = { markers: 0, raw: 0, rawIds: [], lastSeen: 0 };
+let tracking = { markers: 0, raw: 0, rawIds: [], dupes: 0, lastSeen: 0 };
 let netRef = null;
 let sceneLoaded = false;
 let anchorCount = 0;
@@ -68,9 +68,17 @@ function rxTrace(t) { rxLog.push(t); if (rxLog.length > 6) rxLog.shift(); }
 const clockOffsetSamples = [];
 let clockOffset = 0; // serverNow - clientNow
 
-/* Horizontal FOV assumption for the focal estimate; override for a specific device
+/* Field-of-view assumption behind every focal estimate; override per device via
- * via localStorage 'nbx.hfovDeg' (shared by both engines). */
+ * localStorage 'nbx.hfovDeg' (shared by both engines).
+ *
+ * IT APPLIES TO THE LONG AXIS OF THE IMAGE, NOT TO videoWidth. iOS can hand back
+ * a PORTRAIT stream (720x1280), and a phone lens has its wide field across the
+ * sensor's long side either way. Assuming the angle spanned videoWidth made the
+ * focal length 1280/720 = 1.78x too short on portrait streams — the exact factor
+ * measured on device, and the reason ghosts still slid around after the camera
+ * model was "fixed". */
 const HFOV_DEG = parseFloat(localStorage.getItem('nbx.hfovDeg')) || 60;
+const focalPxFor = (w, h) => (Math.max(w, h) / 2) / Math.tan(THREE.MathUtils.degToRad(HFOV_DEG) / 2);
 
 /* Detection cadence. Marker detection (getImageData + ArUco decode over a 960px
  * frame) is by far the most expensive thing per frame, and running it on every
@@ -147,7 +155,7 @@ function updateDbg(fusedQuat, markerCount, extra) {
   dbgPose =
     `eng ${engine}  ${freezeGhosts ? 'FROZEN' : 'moving'}\n` +
     `${modeLine}\n` +
-    `markers ${markerCount}/${tracking.raw} seen [${tracking.rawIds.join(',')}]${extra}\n${rawLine}\nview ${v3(view)}\nup   ${v3(up)}\n` +
+    `markers ${markerCount}/${tracking.raw} seen [${tracking.rawIds.join(',')}]${tracking.dupes ? ' +' + tracking.dupes + 'dup' : ''}${extra}\n${rawLine}\nview ${v3(view)}\nup   ${v3(up)}\n` +
     `pos  ${v3(camera3.position)}`;
   paintDbg();
 }
@@ -346,10 +354,11 @@ function onResize() {
  * geometry project too far from centre — measured at 1.7x on device, which
  * reads as ghosts sliding the wrong way when you tilt.
  *
- * So: take the focal length implied by HFOV_DEG over the full video width, work
+ * So: take the focal length implied by HFOV_DEG over the image's LONG axis (see
- * out how much of the frame survives the cover-crop, and set the fov from that
+ * focalPxFor — the stream may be portrait), work out how much of the frame
- * visible height. Cover crops symmetrically, so the principal point stays
+ * survives the cover-crop, and set the fov from that visible height. Cover crops
- * centred and no lens shift is needed. */
+ * symmetrically, so the principal point stays centred and no lens shift is
+ * needed. */
 const _crop = { x: 1, y: 1 };   // fraction of the video width/height still visible
 function updateCameraIntrinsics() {
   if (!camera3) return;
@@ -357,7 +366,7 @@ function updateCameraIntrinsics() {
   if (!video || !video.videoWidth) { camera3.updateProjectionMatrix(); return; }
 
   const vw = video.videoWidth, vh = video.videoHeight;
-  const focal = (vw / 2) / Math.tan(THREE.MathUtils.degToRad(HFOV_DEG) / 2);
+  const focal = focalPxFor(vw, vh);
   const scale = Math.max(innerWidth / vw, innerHeight / vh);   // object-fit: cover
   const visW = Math.min(vw, innerWidth / scale);
   const visH = Math.min(vh, innerHeight / scale);
@@ -463,20 +472,25 @@ function loop(t) {
 
   if (video && video.readyState >= 2 && (t - lastDetectAt) >= DETECT_INTERVAL_MS) {
     lastDetectAt = t;
-    // downscale for detection speed; corner precision scales with resolution.
+    /* Downscale for detection speed, capping the LONG side. Fixing the WIDTH at
-    const W = 960;
+     * 960 quietly tripled the work on a portrait stream (960x1707 = 1.6M px vs
-    const H = Math.round(W * video.videoHeight / video.videoWidth);
+     * 960x540 = 0.5M), which is most of the cost of a frame. */
-    if (canvas2d.width !== W) { canvas2d.width = W; canvas2d.height = H; }
+    const vw = video.videoWidth, vh = video.videoHeight;
+    const LONG = 960;
+    const W = vw >= vh ? LONG : Math.round(LONG * vw / vh);
+    const H = vw >= vh ? Math.round(LONG * vh / vw) : LONG;
+    if (canvas2d.width !== W || canvas2d.height !== H) { canvas2d.width = W; canvas2d.height = H; }
     ctx2d.drawImage(video, 0, 0, W, H);
     const img = ctx2d.getImageData(0, 0, W, H);
 
-    const focal = W / (2 * Math.tan(THREE.MathUtils.degToRad(HFOV_DEG) / 2));
+    const focal = focalPxFor(W, H);
     if (!poseEst) poseEst = new PoseEstimator(focal);
 
     const markers = detectMarkers(detector, img, fuser.knownIds());
     tracking.markers = markers.length;
     tracking.raw = markers.rawCount || 0;
     tracking.rawIds = markers.rawIds || [];
+    tracking.dupes = markers.dupes || 0;
     if (dbgEl) updateAim(markers, W, H);
     if (markers.length) {
       tracking.lastSeen = performance.now();