merge: precisione rotazione + perf propagate + robustezza server

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

ROADMAP.md

@@ -2,6 +2,36 @@
 
 Lista ragionata di miglioramenti futuri. Priorità = impatto / effort, non urgenza temporale.
 
+## Fase 2 COMPLETATA (precisione rotazione + robustezza + perf)
+
+Root cause della rotazione imprecisa: lo score satura a 1.0 sulla spread
+bitmap dilatata (raggio 4-5) → il refine non vedeva gradiente né in angolo
+né in posizione, e `cv2.minMaxLoc` sul plateau saturo spostava il centro
+sull'angolo della finestra (errore sistematico 3·√2 ≈ 4.24 px).
+
+| Fix | Dettaglio |
+|---|---|
+| Refine su bitmap fine | `_refine_angle` ottimizza su spread raggio 1 (`spread_fine`, cached); score finale ricalcolato su spread coarse per mantenere semantica soglie |
+| Picco sub-pixel nel refine | centroide plateau / fit quadratico al posto di minMaxLoc (bias top-left) |
+| LM least-squares pos+angolo | `_subpixel_refine_lm` riscritto: snap edge ±2px lungo normale + LSQ 3x3 (dx, dy, dθ), ON di default |
+| Round feature offsets | troncamento `astype(int32)` → `np.round` (bias ~0.25 px) |
+| Centro rotazione coerente | `_prepare_padded_template`: rotazione attorno al centro reale del template nel padding (bias ≤0.5 px dipendente dall'angolo) |
+| `_angle_list` include estremo | range parziali ±tol ora testano anche +tol |
+| `_refine_pose_joint` rimosso | Nelder-Mead su funzione a gradini satura: terminava subito; param ora alias di refine_angle |
+| pyramid_propagate di default | kernel windowed (feature campionano l'intera scena: prima il crop troncava le feature → score 0); picchi = massimi locali (non top-K pixel); disattivato automaticamente per template elongati (>2:1) dove il picco top-level non localizza |
+| Piramide 3 livelli default | con clamp automatico sulla dimensione template (min 12 px al top) |
+| Cache scena: hash completo | prima hashava solo i primi 64KB → collisioni tra scene con stessa banda superiore → risultati della scena sbagliata |
+| Web server | lock matcher (race con threadpool FastAPI), LRU `_IMG_CACHE`, clamp ROI ovunque (400/422 invece di 500), `filtro_fp=off` disabilita davvero NCC, `_draw_matches` su crop locale |
+| GUI/legacy | centro overlay `(W-1)/2`→`W/2`, spread_radius default 5→4, EdgeShapeMatcher: angle list endpoint + cap candidati + save template_gray |
+
+Misure (GT sintetica 7 pose, scena 900x700, VPS 2 core):
+- Errore angolare mediano: **2.3° → 0.05°** (step 5°); a step 2° era 4.4° → **0.03°**
+- Errore posizione mediano: **4.24 px → 0.04 px**
+- find GT scene: 4.7s → 1.7s; scena reale 646x482: 1.14s → 0.81s
+- Benchmark suite 16 scenari: 96.5s → 84.2s, match count ≥ baseline
+  (eccezioni: dado_full -1 = match borderline su parte diversa;
+  lama_part_preciso 25→18 con baseline al cap max_matches)
+
 ## Fase 1 COMPLETATA (branch `speedFase1`)
 
 | ID | Voce | Status | Note |
@@ -84,9 +114,14 @@ Benchmark suite 16 scenari (4 immagini × full/part × fast/preciso):
 ## Target performance produzione
 
 Obiettivi da documento tecnico Vision Suite (Fase Beta):
-- [ ] **Precisione posizionale mediana**: <0.5 px → **raggiunto con subpixel (attualmente ~0.1-0.3 px atteso)**
-- [ ] **Precisione angolare mediana**: <1.0° → **raggiunto con refinement (~0.5°)**
-- [ ] **Latency mediana**: <50 ms su 1920×1080 → **attuale ~1.7s su 830×822 (serve GPU o ulteriore CPU)**
+- [x] **Precisione posizionale mediana**: <0.5 px → **0.04 px misurato su GT sintetica (Fase 2)**
+- [x] **Precisione angolare mediana**: <1.0° → **0.05° misurato su GT sintetica (Fase 2)**
+- [ ] **Latency mediana**: <50 ms su 1920×1080 → **~0.8s su 646×482 con 2 core; da misurare su hardware produzione**
 - [ ] **F1 score dataset sintetico**: >0.95 → **da misurare con dataset sintetico**
 
-Prossimo blocker per target: **latency**. Via più promettente: GPU (CuPy) o coarse-to-fine angolare.
+Prossimo blocker per target: **latency**. Nota: i kernel hot sono gia'
+Numba JIT (≈ velocita' C, prange parallelo): un port C++ dei kernel vale
+solo il margine SIMD esplicito (~2-4x con AVX2 su AND+popcount byte-wise).
+Prima di scriverlo conviene esaurire le vie algoritmiche rimaste:
+riduzione varianti al top-level (auto angle step per livello, stile
+Halcon), greediness di default, e GPU (CuPy/OpenCL) per scene 1080p.

benchmarks/test_suite.py

@@ -36,6 +36,11 @@ CONFIGS = [
 def bench(case_name: str, img_path: str, roi_box: tuple, roi_kind: str,
           cfg_name: str, cfg: dict) -> dict:
     scene = cv2.imread(str(TEST_DIR / img_path))
+    if scene is None:
+        # cv2.imread ritorna None silenzioso: senza check il crash arriva
+        # dopo, sullo slice, con un errore criptico.
+        raise FileNotFoundError(
+            f"Immagine di test non trovata o non leggibile: {TEST_DIR / img_path}")
     y0, y1, x0, x1 = roi_box
     roi = scene[y0:y1, x0:x1].copy()
     m = LineShapeMatcher(

pm2d/_jit_kernels.py

@@ -271,6 +271,108 @@ if HAS_NUMBA:
                         acc[y, x] = 0.0
         return acc
 
+    @nb.njit(cache=True, parallel=True, fastmath=True, boundscheck=False)
+    def _jit_score_bitmap_rescored_window(
+        spread: np.ndarray,      # uint8 (H, W) - scena INTERA
+        dx: np.ndarray, dy: np.ndarray, bins: np.ndarray,
+        bit_active: np.uint8,
+        bg: np.ndarray,          # float32 (H, W) - scena intera
+        y0: nb.int64, x0: nb.int64,
+        wh: nb.int64, ww: nb.int64,
+    ) -> np.ndarray:
+        """Score rescored valutato SOLO nella finestra (y0, x0, wh, ww).
+
+        Le feature campionano lo spread dell'intera scena (bounds-checked
+        sui bordi scena): a differenza di chiamare il kernel su un crop,
+        le feature che escono dalla finestra NON contano come miss.
+        Usato dal path pyramid_propagate: costo ∝ area finestra.
+        """
+        H, W = spread.shape
+        N = dx.shape[0]
+        acc = np.zeros((wh, ww), dtype=np.float32)
+        for yi in nb.prange(wh):
+            y = y0 + yi
+            for i in range(N):
+                b = bins[i]
+                mask = np.uint8(1) << b
+                if (bit_active & mask) == 0:
+                    continue
+                yy = y + dy[i]
+                if yy < 0 or yy >= H:
+                    continue
+                ddx = dx[i]
+                xi_lo = 0
+                xi_hi = ww
+                lo = -(x0 + ddx)
+                if lo > xi_lo:
+                    xi_lo = lo
+                hi = W - (x0 + ddx)
+                if hi < xi_hi:
+                    xi_hi = hi
+                for xi in range(xi_lo, xi_hi):
+                    if spread[yy, x0 + xi + ddx] & mask:
+                        acc[yi, xi] += 1.0
+        if N > 0:
+            inv = 1.0 / N
+            for yi in nb.prange(wh):
+                for xi in range(ww):
+                    v = acc[yi, xi] * inv
+                    bgv = bg[y0 + yi, x0 + xi]
+                    if bgv < 1.0:
+                        r = (v - bgv) / (1.0 - bgv + 1e-6)
+                        acc[yi, xi] = r if r > 0.0 else 0.0
+                    else:
+                        acc[yi, xi] = 0.0
+        return acc
+
+    @nb.njit(cache=True, parallel=True, fastmath=True, boundscheck=False)
+    def _jit_score_bitmap_rescored_window_u16(
+        spread: np.ndarray,      # uint16 (H, W)
+        dx: np.ndarray, dy: np.ndarray, bins: np.ndarray,
+        bit_active: np.uint16,
+        bg: np.ndarray,
+        y0: nb.int64, x0: nb.int64,
+        wh: nb.int64, ww: nb.int64,
+    ) -> np.ndarray:
+        """Versione uint16 (polarity 16-bin) del kernel windowed."""
+        H, W = spread.shape
+        N = dx.shape[0]
+        acc = np.zeros((wh, ww), dtype=np.float32)
+        for yi in nb.prange(wh):
+            y = y0 + yi
+            for i in range(N):
+                b = bins[i]
+                mask = np.uint16(1) << b
+                if (bit_active & mask) == 0:
+                    continue
+                yy = y + dy[i]
+                if yy < 0 or yy >= H:
+                    continue
+                ddx = dx[i]
+                xi_lo = 0
+                xi_hi = ww
+                lo = -(x0 + ddx)
+                if lo > xi_lo:
+                    xi_lo = lo
+                hi = W - (x0 + ddx)
+                if hi < xi_hi:
+                    xi_hi = hi
+                for xi in range(xi_lo, xi_hi):
+                    if spread[yy, x0 + xi + ddx] & mask:
+                        acc[yi, xi] += 1.0
+        if N > 0:
+            inv = 1.0 / N
+            for yi in nb.prange(wh):
+                for xi in range(ww):
+                    v = acc[yi, xi] * inv
+                    bgv = bg[y0 + yi, x0 + xi]
+                    if bgv < 1.0:
+                        r = (v - bgv) / (1.0 - bgv + 1e-6)
+                        acc[yi, xi] = r if r > 0.0 else 0.0
+                    else:
+                        acc[yi, xi] = 0.0
+        return acc
+
     @nb.njit(cache=True, parallel=True, fastmath=True, boundscheck=False)
     def _jit_top_max_per_variant(
         spread: np.ndarray,            # uint8 (H, W)
@@ -426,6 +528,9 @@ if HAS_NUMBA:
         _jit_top_max_per_variant(
             spread, dx, dy, b, offsets, np.uint8(0xFF), bg_pv, scale_idx,
         )
+        _jit_score_bitmap_rescored_window(
+            spread, dx, dy, b, np.uint8(0xFF), bg, 4, 4, 8, 8,
+        )
         _jit_popcount_density(spread)
         spread16 = np.zeros((32, 32), dtype=np.uint16)
         _jit_score_bitmap_rescored_u16(
@@ -447,6 +552,12 @@ else:  # pragma: no cover
     def _jit_score_bitmap_rescored_strided(spread, dx, dy, bins, bit_active, bg, stride):
         raise RuntimeError("numba non disponibile")
 
+    def _jit_score_bitmap_rescored_window(spread, dx, dy, bins, bit_active, bg, y0, x0, wh, ww):
+        raise RuntimeError("numba non disponibile")
+
+    def _jit_score_bitmap_rescored_window_u16(spread, dx, dy, bins, bit_active, bg, y0, x0, wh, ww):
+        raise RuntimeError("numba non disponibile")
+
     def _jit_score_bitmap_greedy(spread, dx, dy, bins, bit_active, min_score, greediness):
         raise RuntimeError("numba non disponibile")
 
@@ -524,6 +635,39 @@ def score_bitmap_rescored(
     return np.maximum(0.0, out).astype(np.float32)
 
 
+def score_bitmap_rescored_window(
+    spread: np.ndarray, dx: np.ndarray, dy: np.ndarray, bins: np.ndarray,
+    bit_active: int, bg: np.ndarray,
+    y0: int, x0: int, wh: int, ww: int,
+) -> np.ndarray:
+    """Score rescored solo nella finestra (y0, x0, wh, ww) della scena.
+
+    Le feature campionano l'INTERA scena: feature fuori finestra ma dentro
+    scena contano correttamente (chiamare il kernel su un crop le tratta
+    come miss e azzera lo score — il bug che rendeva inutilizzabile il
+    path pyramid_propagate). Fallback no-numba: kernel pieno + slice.
+    """
+    if HAS_NUMBA and len(dx) > 0:
+        dx_c = np.ascontiguousarray(dx, dtype=np.int32)
+        dy_c = np.ascontiguousarray(dy, dtype=np.int32)
+        bins_c = np.ascontiguousarray(bins, dtype=np.int8)
+        bg_c = np.ascontiguousarray(bg, dtype=np.float32)
+        if spread.dtype == np.uint16:
+            return _jit_score_bitmap_rescored_window_u16(
+                np.ascontiguousarray(spread, dtype=np.uint16),
+                dx_c, dy_c, bins_c, np.uint16(bit_active), bg_c,
+                int(y0), int(x0), int(wh), int(ww),
+            )
+        return _jit_score_bitmap_rescored_window(
+            np.ascontiguousarray(spread, dtype=np.uint8),
+            dx_c, dy_c, bins_c, np.uint8(bit_active), bg_c,
+            int(y0), int(x0), int(wh), int(ww),
+        )
+    # Fallback (lento, solo senza numba): score full-frame + slice finestra
+    full = score_bitmap_rescored(spread, dx, dy, bins, bit_active, bg)
+    return full[y0:y0 + wh, x0:x0 + ww]
+
+
 def score_bitmap_greedy(
     spread: np.ndarray, dx: np.ndarray, dy: np.ndarray, bins: np.ndarray,
     bit_active: int, min_score: float, greediness: float,

pm2d/auto_tune.py

@@ -61,6 +61,8 @@ def detect_rotational_symmetry(
 
     center = (w / 2.0, h / 2.0)
     ref = mag
+    # ref è costante nel loop sugli angoli: centra una volta sola
+    rm = ref - ref.mean()
 
     correlations: list[tuple[float, float]] = []
     for ang in np.arange(step_deg, 360.0, step_deg):
@@ -68,7 +70,6 @@ def detect_rotational_symmetry(
         rot = cv2.warpAffine(
             mag, M, (w, h), borderValue=0.0,
         )
-        rm = ref - ref.mean()
         rs = rot - rot.mean()
         denom = np.sqrt((rm * rm).sum() * (rs * rs).sum()) + 1e-9
         c = float((rm * rs).sum() / denom)

pm2d/bench.py

@@ -0,0 +1,179 @@
+"""Benchmark suite per LineShapeMatcher.
+
+Usage:
+    python -m pm2d.bench [--quick]
+
+Misura tempi find() su 3 template-tipo × 3 scene-tipo × N config:
+- Template: rettangolo 80×80, L-shape 120×120, cerchio 150×150
+- Scene: pulita 800×600, cluttered 1080×1920, multi-pezzo 1080×1920
+- Config: baseline, polarity, gpu, pyramid_propagate, greediness=0.7
+
+Per ogni config stampa: ms/find, ms per fase (profile), n. match.
+Output tabellare per detectare regressioni in CI.
+"""
+
+from __future__ import annotations
+
+import argparse
+import time
+
+import cv2
+import numpy as np
+
+from pm2d.line_matcher import LineShapeMatcher, opencl_available
+
+
+# ---------- Sintetizzatori template/scena ----------
+
+def _tpl_rect() -> np.ndarray:
+    t = np.zeros((80, 80, 3), np.uint8)
+    cv2.rectangle(t, (15, 15), (65, 65), (255, 255, 255), 3)
+    return t
+
+
+def _tpl_lshape() -> np.ndarray:
+    t = np.zeros((120, 120, 3), np.uint8)
+    cv2.rectangle(t, (20, 20), (50, 100), (255, 255, 255), -1)
+    cv2.rectangle(t, (20, 70), (100, 100), (255, 255, 255), -1)
+    return t
+
+
+def _tpl_circle() -> np.ndarray:
+    t = np.zeros((150, 150, 3), np.uint8)
+    cv2.circle(t, (75, 75), 60, (255, 255, 255), 4)
+    return t
+
+
+def _scene_clean(W: int, H: int, n_pieces: int = 1) -> np.ndarray:
+    np.random.seed(0)
+    s = np.zeros((H, W, 3), np.uint8)
+    for _ in range(n_pieces):
+        cx = np.random.randint(80, W - 80)
+        cy = np.random.randint(80, H - 80)
+        cv2.rectangle(s, (cx - 25, cy - 25), (cx + 25, cy + 25), (255, 255, 255), 3)
+    return s
+
+
+def _scene_cluttered(W: int, H: int) -> np.ndarray:
+    np.random.seed(0)
+    s = np.random.randint(50, 200, (H, W, 3), np.uint8)
+    cv2.rectangle(s, (300, 200), (350, 250), (255, 255, 255), 3)
+    cv2.rectangle(s, (1500, 800), (1550, 850), (255, 255, 255), 3)
+    return s
+
+
+# ---------- Single benchmark ----------
+
+def _bench_config(template, scene, config_name: str,
+                  init_kw: dict, find_kw: dict,
+                  n_iter: int = 5) -> dict:
+    m = LineShapeMatcher(**init_kw)
+    t0 = time.perf_counter()
+    n_var = m.train(template)
+    t_train = time.perf_counter() - t0
+
+    # Warmup (Numba JIT)
+    m.find(scene, **find_kw)
+    m.find(scene, **find_kw)
+
+    # Run
+    times_ms = []
+    for _ in range(n_iter):
+        t0 = time.perf_counter()
+        matches = m.find(scene, **find_kw)
+        times_ms.append((time.perf_counter() - t0) * 1000.0)
+
+    # Profile (1 iter)
+    m.find(scene, profile=True, **find_kw)
+    prof = m.get_last_profile() or {}
+
+    return {
+        "config": config_name,
+        "n_variants": n_var,
+        "t_train_s": round(t_train, 3),
+        "ms_avg": round(float(np.mean(times_ms)), 1),
+        "ms_min": round(float(np.min(times_ms)), 1),
+        "ms_max": round(float(np.max(times_ms)), 1),
+        "n_matches": len(matches),
+        "profile_ms": {k: round(v, 1) for k, v in prof.items()},
+    }
+
+
+# ---------- Suite ----------
+
+CONFIGS = [
+    ("baseline",
+     {"angle_step_deg": 10, "pyramid_levels": 2},
+     {"min_score": 0.4, "verify_threshold": 0.2}),
+    ("polarity",
+     {"angle_step_deg": 10, "pyramid_levels": 2, "use_polarity": True},
+     {"min_score": 0.4, "verify_threshold": 0.2}),
+    ("propagate",
+     {"angle_step_deg": 10, "pyramid_levels": 3},
+     {"min_score": 0.4, "verify_threshold": 0.2,
+      "pyramid_propagate": True, "propagate_topk": 4}),
+    ("greedy_07",
+     {"angle_step_deg": 10, "pyramid_levels": 2},
+     {"min_score": 0.4, "verify_threshold": 0.2, "greediness": 0.7}),
+    ("stride2",
+     {"angle_step_deg": 10, "pyramid_levels": 2},
+     {"min_score": 0.4, "verify_threshold": 0.2, "coarse_stride": 2}),
+]
+
+if opencl_available():
+    CONFIGS.append(
+        ("gpu_umat",
+         {"angle_step_deg": 10, "pyramid_levels": 2, "use_gpu": True},
+         {"min_score": 0.4, "verify_threshold": 0.2})
+    )
+
+
+SCENARIOS = [
+    ("rect_80 vs scene_800x600", _tpl_rect, lambda: _scene_clean(800, 600, 1)),
+    ("lshape_120 vs scene_1080x1920_clutter",
+     _tpl_lshape, lambda: _scene_cluttered(1920, 1080)),
+    ("circle_150 vs scene_clean_3pieces",
+     _tpl_circle, lambda: _scene_clean(1920, 1080, 3)),
+]
+
+
+def run(quick: bool = False) -> int:
+    n_iter = 2 if quick else 5
+    print(f"=== PM2D Benchmark Suite ({len(SCENARIOS)} scenarios x "
+          f"{len(CONFIGS)} configs, n_iter={n_iter}) ===\n")
+    rows = []
+    for sc_name, tpl_fn, scn_fn in SCENARIOS:
+        template = tpl_fn()
+        scene = scn_fn()
+        print(f"--- Scenario: {sc_name} (tpl={template.shape}, "
+              f"scn={scene.shape}) ---")
+        for cfg_name, init_kw, find_kw in CONFIGS:
+            r = _bench_config(template, scene, cfg_name, init_kw, find_kw,
+                              n_iter=n_iter)
+            r["scenario"] = sc_name
+            rows.append(r)
+            prof_str = " ".join(
+                f"{k}={v:.1f}" for k, v in r["profile_ms"].items()
+            )
+            print(f"  {cfg_name:14s}  {r['ms_avg']:6.1f}ms  "
+                  f"(min {r['ms_min']:.1f} max {r['ms_max']:.1f})  "
+                  f"vars={r['n_variants']:3d}  "
+                  f"matches={r['n_matches']:2d}")
+            if prof_str:
+                print(f"     profile: {prof_str}")
+        print()
+    print("=== Done ===")
+    return 0
+
+
+def main(argv: list[str] | None = None) -> int:
+    p = argparse.ArgumentParser(description="PM2D benchmark suite")
+    p.add_argument("--quick", action="store_true",
+                   help="2 iterazioni per config invece di 5 (smoke test)")
+    args = p.parse_args(argv)
+    return run(quick=args.quick)
+
+
+if __name__ == "__main__":
+    import sys
+    sys.exit(main())

pm2d/gui.py

@@ -196,8 +196,10 @@ def _warp_template_edges_to_scene(
     edge = cv2.Canny(template_gray, canny_low, canny_high)
     # Matrice affine: scala + rotazione attorno al centro template, poi traslazione
     Ht, Wt = h, w
-    cx_t = (Wt - 1) / 2.0
-    cy_t = (Ht - 1) / 2.0
+    # Centro coerente con la convenzione train (center = w / 2.0, no -1):
+    # (Wt-1)/2 introduceva uno shift di 0.5px per template di lato pari.
+    cx_t = Wt / 2.0
+    cy_t = Ht / 2.0
     M = cv2.getRotationMatrix2D((cx_t, cy_t), angle_deg, scale)
     # Traslazione per portare centro template a (cx, cy) della scena
     M[0, 2] += cx - cx_t
@@ -492,7 +494,9 @@ def run(
     num_features: int = 96,
     weak_grad: float = 30.0,
     strong_grad: float = 60.0,
-    spread_radius: int = 5,
+    # 4 allineato col default del matcher: raggio 5 peggiora la precisione
+    # di rotazione (spread troppo largo appiattisce il picco angolare).
+    spread_radius: int = 4,
     pyramid_levels: int = 3,
     min_score: float = 0.55,
     max_matches: int = 25,

pm2d/matcher.py

@@ -91,8 +91,16 @@ class EdgeShapeMatcher:
         a0, a1 = self.angle_range_deg
         if self.angle_step_deg <= 0 or a0 >= a1:
             return [float(a0)]
-        n = int(np.floor((a1 - a0) / self.angle_step_deg))
-        return [float(a0 + i * self.angle_step_deg) for i in range(n)]
+        # n+1 valori per includere l'estremo superiore del range: con il
+        # solo floor un range [0, 90] step 5 si fermava a 85° (off-by-one).
+        n = int(np.floor((a1 - a0) / self.angle_step_deg)) + 1
+        angles = [float(a0 + i * self.angle_step_deg) for i in range(n)]
+        if a1 - a0 >= 360.0:
+            # Range che copre il giro completo: a0+360° è la stessa pose di
+            # a0, escludi il duplicato (variante inutile in train/find).
+            eps = 1e-6
+            angles = [a for a in angles if a < a0 + 360.0 - eps]
+        return angles
 
     def train(self, template_bgr: np.ndarray) -> int:
         """Genera varianti per tutte le combinazioni (angolo, scala)."""
@@ -222,6 +230,14 @@ class EdgeShapeMatcher:
             for y, x in zip(ys, xs):
                 candidates.append((float(res[y, x]), int(x), int(y), ti))
 
+        # Cap candidati top-level: senza limite np.where con soglia bassa
+        # può generare migliaia di candidati, ognuno con un matchTemplate
+        # full-res nel refinement. Tieni solo i migliori per score.
+        max_candidates = max(1, max_matches * 10)
+        if len(candidates) > max_candidates:
+            candidates.sort(key=lambda c: -c[0])
+            candidates = candidates[:max_candidates]
+
         # Refinement a risoluzione piena: per ogni candidato top, finestra locale
         refined: list[tuple[float, int, int, int]] = []
         margin = sf + 4
@@ -294,6 +310,10 @@ class EdgeShapeMatcher:
         )
         arrays = {f"edge_{i}": t.edge for i, t in enumerate(self.templates)}
         arrays.update({f"mask_{i}": t.mask for i, t in enumerate(self.templates)})
+        # Persisti anche il grayscale originale: senza, l'overlay edge
+        # spariva dopo load() (template_gray restava None).
+        if self.template_gray is not None:
+            arrays["template_gray"] = self.template_gray
         np.savez_compressed(path, params=params, meta=meta, **arrays)
 
     @classmethod
@@ -312,6 +332,10 @@ class EdgeShapeMatcher:
             top_score_factor=float(p[12]) if len(p) > 12 else 0.6,
         )
         m.template_size = (int(p[8]), int(p[9]))
+        # Retrocompatibilità: modelli salvati prima non hanno template_gray
+        # (resta None: overlay edge non disponibile ma find() funziona).
+        if "template_gray" in z.files:
+            m.template_gray = z["template_gray"]
         meta = z["meta"]
         for i in range(len(meta)):
             m.templates.append(

pm2d/web/server.py

@@ -12,6 +12,7 @@ from __future__ import annotations
 import hashlib
 import os
 import tempfile
+import threading
 import time
 import uuid
 from collections import OrderedDict
@@ -64,14 +65,23 @@ STATIC_DIR.mkdir(exist_ok=True)
 CACHE_DIR = Path(tempfile.gettempdir()) / "pm2d_cache"
 CACHE_DIR.mkdir(exist_ok=True)
 
-# Cache in-memory (soft, ricaricata da disco se mancante)
-_IMG_CACHE: dict[str, np.ndarray] = {}
+# Cache in-memory (soft, ricaricata da disco se mancante).
+# LRU con capacità limitata: senza eviction le immagini si accumulavano
+# senza limite (leak di memoria su server long-running).
+_IMG_CACHE: OrderedDict[str, np.ndarray] = OrderedDict()
+_IMG_CACHE_SIZE = 64
 
 # Cache matcher addestrati: (roi_hash, params_hash) -> LineShapeMatcher
 # LRU con capacità limitata
 _MATCHER_CACHE: OrderedDict = OrderedDict()
 _MATCHER_CACHE_SIZE = 8
 
+# Lock globale matcher: gli endpoint girano nel threadpool FastAPI ma i
+# matcher condivisi (_MATCHER_CACHE, _RECIPE_MATCHERS) mutano stato interno
+# durante train()/find(). Serializzare il matching è la soluzione semplice
+# e corretta (un lock per-ricetta sarebbe over-engineering).
+_MATCHER_LOCK = threading.Lock()
+
 
 def _matcher_cache_key(roi: np.ndarray, tech: dict) -> str:
     h = hashlib.md5()
@@ -102,23 +112,32 @@ def _cache_put_matcher(key: str, matcher) -> None:
         _MATCHER_CACHE.popitem(last=False)
 
 
+def _img_cache_put(key: str, value: np.ndarray) -> None:
+    """Inserisce in _IMG_CACHE con eviction LRU (cap _IMG_CACHE_SIZE)."""
+    _IMG_CACHE[key] = value
+    _IMG_CACHE.move_to_end(key)
+    while len(_IMG_CACHE) > _IMG_CACHE_SIZE:
+        _IMG_CACHE.popitem(last=False)
+
+
 def _store_image(img: np.ndarray) -> str:
     iid = uuid.uuid4().hex[:12]
     cv2.imwrite(str(CACHE_DIR / f"{iid}.png"), img)
-    _IMG_CACHE[iid] = img
+    _img_cache_put(iid, img)
     return iid
 
 
 def _load_image(iid: str) -> np.ndarray | None:
     cached = _IMG_CACHE.get(iid)
     if cached is not None:
+        _IMG_CACHE.move_to_end(iid)  # LRU touch
         return cached
     p = CACHE_DIR / f"{iid}.png"
     if not p.exists():
         return None
     img = cv2.imread(str(p))
     if img is not None:
-        _IMG_CACHE[iid] = img
+        _img_cache_put(iid, img)
     return img
 
 app = FastAPI(title="PM2D Webapp", version="1.0.0")
@@ -131,6 +150,39 @@ def _encode_png(img: np.ndarray) -> bytes:
     return buf.tobytes()
 
 
+def _clamp_roi(x: int, y: int, w: int, h: int,
+               img_w: int, img_h: int) -> tuple[int, int, int, int]:
+    """Clampa la ROI dentro i limiti immagine.
+
+    Una ROI fuori immagine causava slice vuote → crash 500 negli endpoint
+    che non clampavano. Solleva 400 se la ROI risultante è degenere
+    (lato < 16 px: sotto questa soglia il train non estrae abbastanza
+    edge feature e produce 0 varianti → find() esplode con 500).
+    """
+    x = max(0, min(int(x), img_w - 1))
+    y = max(0, min(int(y), img_h - 1))
+    w = min(int(w), img_w - x)
+    h = min(int(h), img_h - y)
+    if w < 16 or h < 16:
+        raise HTTPException(
+            400, f"ROI fuori immagine o degenere: [{x}, {y}, {w}, {h}] "
+                 f"su immagine {img_w}x{img_h} (lato minimo 16 px)")
+    return x, y, w, h
+
+
+def _check_trained(m: "LineShapeMatcher", n_variants: int) -> None:
+    """Solleva 422 se il train non ha prodotto varianti.
+
+    Succede con ROI senza contrasto (sfondo uniforme) o troppo piccola:
+    senza questo check il find() successivo esplode con RuntimeError → 500.
+    """
+    if n_variants <= 0 or not m.variants:
+        raise HTTPException(
+            422, "La ROI non contiene abbastanza edge feature per il "
+                 "training (zona troppo uniforme o piccola): scegliere "
+                 "una regione con contorni netti")
+
+
 def _draw_matches(scene: np.ndarray, matches: list[Match],
                    template_gray: np.ndarray | None,
                    matcher: "LineShapeMatcher | None" = None) -> np.ndarray:
@@ -143,17 +195,6 @@ def _draw_matches(scene: np.ndarray, matches: list[Match],
     nell'anteprima modello.
     """
     out = scene.copy()
-    # Baricentro UCS in coord template (calcolato una volta dal matcher
-    # se disponibile): mean delle feature di una variante a 0°. Questo e'
-    # lo stesso baricentro mostrato nell'anteprima modello.
-    bary_dx = bary_dy = 0.0
-    if matcher is not None and matcher.variants:
-        # Trova variante con angle_deg piu vicino a 0
-        v0 = min(matcher.variants, key=lambda v: abs(v.angle_deg))
-        if len(v0.levels[0].dx) > 0:
-            bary_dx = float(np.mean(v0.levels[0].dx))
-            bary_dy = float(np.mean(v0.levels[0].dy))
-
     # Lunghezza assi UCS: stessa formula dell'anteprima modello
     # (0.15 * max lato template) scalata per m.scale → coerenza dimensionale.
     if matcher is not None and matcher.template_size != (0, 0):
@@ -161,38 +202,75 @@ def _draw_matches(scene: np.ndarray, matches: list[Match],
     else:
         L_base = 30
     H_scene, W_scene = scene.shape[:2]
+
     for i, m in enumerate(matches):
-        # Proietta baricentro template alla pose del match.
-        # cv2.getRotationMatrix2D con angle positivo applica:
-        #   new_x = cos*x + sin*y    new_y = -sin*x + cos*y
-        # Visivamente in image y-down e' rotazione anti-clockwise.
+        # UCS posizionato esattamente sul CENTRO POSE del match (m.cx, m.cy):
+        # equivale al centro template traslato alla scena, ruotato con
+        # m.angle_deg. Coerente con UCS dell'anteprima modello che ora
+        # e' anche sul centro ROI (vedi preview_edges).
         ax = np.deg2rad(m.angle_deg)
         ca, sa = np.cos(ax), np.sin(ax)
-        bx_scene = m.cx + (bary_dx * ca + bary_dy * sa) * m.scale
-        by_scene = m.cy + (-bary_dx * sa + bary_dy * ca) * m.scale
-        cx, cy = int(round(bx_scene)), int(round(by_scene))
+        cx, cy = int(round(m.cx)), int(round(m.cy))
         # Overlay edge del modello orientato (richiesta utente):
         # warpa template alla pose, applica hysteresis identica al matcher,
-        # disegna pixel edge come overlay verde tenue.
+        # disegna pixel edge come overlay verde tenue. Maschera col
+        # _train_mask warpato + erode per rimuovere edge sui BORDI del
+        # rettangolo template (transizione bordo nero → scena = falso edge
+        # che appariva come "ROI" attorno a ogni match).
         if template_gray is not None and matcher is not None:
             t = template_gray
             th, tw = t.shape
-            cx_t = (tw - 1) / 2.0; cy_t = (th - 1) / 2.0
-            M = cv2.getRotationMatrix2D((cx_t, cy_t), m.angle_deg, m.scale)
-            M[0, 2] += m.cx - cx_t
-            M[1, 2] += m.cy - cy_t
-            warped_gray = cv2.warpAffine(
-                t, M, (W_scene, H_scene),
-                flags=cv2.INTER_LINEAR, borderValue=0)
-            mag, _ = matcher._gradient(warped_gray)
-            if matcher.weak_grad < matcher.strong_grad:
-                edge_mask = matcher._hysteresis_mask(mag)
-            else:
-                edge_mask = mag >= matcher.strong_grad
-            if edge_mask.any():
-                edge_overlay = np.zeros_like(out)
-                edge_overlay[edge_mask] = (0, 220, 0)  # verde brillante
-                out = cv2.addWeighted(out, 1.0, edge_overlay, 0.6, 0)
+            # Centro template coerente col training: in train si usa
+            # `center = (diag / 2.0, diag / 2.0)` (no -1). Usare (tw-1)/2
+            # introduceva uno shift di 0.5px per template di lato pari.
+            cx_t = tw / 2.0; cy_t = th / 2.0
+            # Lavora su un CROP locale della scena di lato = diagonale del
+            # template ruotato+scalato (+margine), come _verify_ncc: warp
+            # + Sobel sull'INTERA scena per ogni match erano O(W·H) cadauno
+            # (costosissimo su scene grandi con molti match).
+            diag = int(np.ceil(np.hypot(tw, th) * m.scale)) + 8
+            x0 = int(round(m.cx)) - diag // 2
+            y0 = int(round(m.cy)) - diag // 2
+            gx0 = max(0, x0); gy0 = max(0, y0)
+            gx1 = min(W_scene, x0 + diag); gy1 = min(H_scene, y0 + diag)
+            cw, ch_ = gx1 - gx0, gy1 - gy0
+            if cw >= 3 and ch_ >= 3:
+                M = cv2.getRotationMatrix2D((cx_t, cy_t), m.angle_deg, m.scale)
+                # Porta il centro template a (m.cx - gx0, m.cy - gy0) del crop
+                M[0, 2] += (m.cx - gx0) - cx_t
+                M[1, 2] += (m.cy - gy0) - cy_t
+                warped_gray = cv2.warpAffine(
+                    t, M, (cw, ch_),
+                    flags=cv2.INTER_LINEAR, borderValue=0)
+                # Maschera: train_mask se disponibile, altrimenti rettangolo pieno
+                mask_src = (matcher._train_mask if matcher._train_mask is not None
+                            else np.full((th, tw), 255, dtype=np.uint8))
+                warped_mask = cv2.warpAffine(
+                    mask_src, M, (cw, ch_),
+                    flags=cv2.INTER_NEAREST, borderValue=0)
+                # Erode minimo (3x3) per togliere SOLO artefatti border-padding
+                # (~1px di bordo nero da warpAffine borderValue=0). Erode piu'
+                # grande spostava visualmente l'edge verso l'interno e creava
+                # apparente "traslazione fissa" rispetto al bordo del pezzo.
+                kernel_er = np.ones((3, 3), np.uint8)
+                warped_mask = cv2.erode(warped_mask, kernel_er)
+                mag, _ = matcher._gradient(warped_gray)
+                if matcher.weak_grad < matcher.strong_grad:
+                    edge_mask = matcher._hysteresis_mask(mag)
+                else:
+                    edge_mask = mag >= matcher.strong_grad
+                edge_mask = edge_mask & (warped_mask > 0)
+                if edge_mask.any():
+                    # Edge ritraslati nel sistema scena: blend solo sul crop
+                    # (addWeighted lascia invariati i pixel con overlay nullo,
+                    # quindi l'output visivo è identico al full-frame).
+                    sub = out[gy0:gy1, gx0:gx1]
+                    edge_overlay = np.zeros_like(sub)
+                    # Ciano (cambiato da verde): non collide col verde dell'asse
+                    # Y dell'UCS che altrimenti scompariva nell'overlay edge.
+                    edge_overlay[edge_mask] = (255, 200, 0)  # ciano (BGR)
+                    out[gy0:gy1, gx0:gx1] = cv2.addWeighted(
+                        sub, 1.0, edge_overlay, 0.6, 0)
         L = max(20, int(L_base * m.scale))
         # X axis = rotazione di (1, 0) con cv2 matrix → (cos, -sin)
         x_end = (int(cx + L * ca), int(cy - L * sa))
@@ -237,7 +315,9 @@ class MatchParams(BaseModel):
     num_features: int = 96
     weak_grad: float = 30.0
     strong_grad: float = 60.0
-    spread_radius: int = 5
+    # 4 allineato col default del matcher: raggio 5 peggiora la precisione
+    # di rotazione (spread troppo largo appiattisce il picco angolare).
+    spread_radius: int = 4
     pyramid_levels: int = 3
     verify_threshold: float = 0.4
 
@@ -398,7 +478,13 @@ def _simple_to_technical(
         "min_score": p.min_score,
         "max_matches": p.max_matches,
         "nms_radius": 0,
-        "verify_threshold": FILTRO_FP_MAP.get(p.filtro_fp, 0.35),
+        # Fallback = livello "medio" della mappa (no valore hardcoded
+        # che divergerebbe se la mappa cambia).
+        "verify_threshold": FILTRO_FP_MAP.get(p.filtro_fp, FILTRO_FP_MAP["medio"]),
+        # "off" deve disabilitare DAVVERO il verify NCC: passare solo
+        # verify_threshold=0.0 lascerebbe attivo il calcolo NCC (che può
+        # comunque scartare match con score negativo / patch uniformi).
+        "verify_ncc": p.filtro_fp != "off",
         "scale_penalty": p.penalita_scala,
     }
 
@@ -532,9 +618,7 @@ def match(p: MatchParams):
     if model is None or scene is None:
         raise HTTPException(404, "Immagini non trovate")
     x, y, w, h = p.roi
-    x = max(0, x); y = max(0, y)
-    w = max(1, min(w, model.shape[1] - x))
-    h = max(1, min(h, model.shape[0] - y))
+    x, y, w, h = _clamp_roi(x, y, w, h, model.shape[1], model.shape[0])
     roi_img = model[y:y + h, x:x + w]
 
     tech_for_cache = {
@@ -548,33 +632,38 @@ def match(p: MatchParams):
         "pyramid_levels": p.pyramid_levels,
     }
     key = _matcher_cache_key(roi_img, tech_for_cache)
-    m = _cache_get_matcher(key)
-    if m is None:
-        m = LineShapeMatcher(
-            num_features=p.num_features,
-            weak_grad=p.weak_grad, strong_grad=p.strong_grad,
-            angle_range_deg=(p.angle_min, p.angle_max),
-            angle_step_deg=p.angle_step,
-            scale_range=(p.scale_min, p.scale_max),
-            scale_step=p.scale_step,
-            spread_radius=p.spread_radius,
-            pyramid_levels=p.pyramid_levels,
+    # Lock globale: matcher condivisi tra thread del pool FastAPI
+    with _MATCHER_LOCK:
+        m = _cache_get_matcher(key)
+        if m is None:
+            m = LineShapeMatcher(
+                num_features=p.num_features,
+                weak_grad=p.weak_grad, strong_grad=p.strong_grad,
+                angle_range_deg=(p.angle_min, p.angle_max),
+                angle_step_deg=p.angle_step,
+                scale_range=(p.scale_min, p.scale_max),
+                scale_step=p.scale_step,
+                spread_radius=p.spread_radius,
+                pyramid_levels=p.pyramid_levels,
+            )
+            t0 = time.time(); n = m.train(roi_img); t_train = time.time() - t0
+            _check_trained(m, n)
+            _cache_put_matcher(key, m)
+        else:
+            n = len(m.variants); t_train = 0.0
+        nms = p.nms_radius if p.nms_radius > 0 else None
+        t0 = time.time()
+        matches = m.find(
+            scene, min_score=p.min_score, max_matches=p.max_matches,
+            nms_radius=nms, verify_threshold=p.verify_threshold,
+            # Soglia 0 = filtro FP disattivato: skippa proprio il calcolo NCC
+            verify_ncc=p.verify_threshold > 0.0,
         )
-        t0 = time.time(); n = m.train(roi_img); t_train = time.time() - t0
-        _cache_put_matcher(key, m)
-    else:
-        n = len(m.variants); t_train = 0.0
-    nms = p.nms_radius if p.nms_radius > 0 else None
-    t0 = time.time()
-    matches = m.find(
-        scene, min_score=p.min_score, max_matches=p.max_matches,
-        nms_radius=nms, verify_threshold=p.verify_threshold,
-    )
-    t_find = time.time() - t0
+        t_find = time.time() - t0
 
-    # Render annotated image
-    tg = cv2.cvtColor(roi_img, cv2.COLOR_BGR2GRAY)
-    annotated = _draw_matches(scene, matches, tg, matcher=m)
+        # Render annotated image
+        tg = cv2.cvtColor(roi_img, cv2.COLOR_BGR2GRAY)
+        annotated = _draw_matches(scene, matches, tg, matcher=m)
     ann_id = _store_image(annotated)
 
     return MatchResp(
@@ -601,9 +690,7 @@ def match_simple(p: SimpleMatchParams):
     if model is None or scene is None:
         raise HTTPException(404, "Immagini non trovate")
     x, y, w, h = p.roi
-    x = max(0, x); y = max(0, y)
-    w = max(1, min(w, model.shape[1] - x))
-    h = max(1, min(h, model.shape[0] - y))
+    x, y, w, h = _clamp_roi(x, y, w, h, model.shape[1], model.shape[0])
     roi_img = model[y:y + h, x:x + w]
 
     tech = _simple_to_technical(p, roi_img)
@@ -612,47 +699,52 @@ def match_simple(p: SimpleMatchParams):
     # Halcon-mode init params: incidono sul training, includere in cache key
     halcon_init_key = f"|pol={p.use_polarity}|gpu={p.use_gpu}"
     key = key + halcon_init_key
-    m = _cache_get_matcher(key)
-    if m is None:
-        m = LineShapeMatcher(
-            num_features=tech["num_features"],
-            weak_grad=tech["weak_grad"], strong_grad=tech["strong_grad"],
-            angle_range_deg=(tech["angle_min"], tech["angle_max"]),
-            angle_step_deg=tech["angle_step"],
-            scale_range=(tech["scale_min"], tech["scale_max"]),
-            scale_step=tech["scale_step"],
-            spread_radius=tech["spread_radius"],
-            min_feature_spacing=tech.get("min_feature_spacing", 3),
-            pyramid_levels=tech["pyramid_levels"],
-            use_polarity=p.use_polarity,
-            use_gpu=p.use_gpu,
+    # Lock globale: matcher condivisi tra thread del pool FastAPI
+    with _MATCHER_LOCK:
+        m = _cache_get_matcher(key)
+        if m is None:
+            m = LineShapeMatcher(
+                num_features=tech["num_features"],
+                weak_grad=tech["weak_grad"], strong_grad=tech["strong_grad"],
+                angle_range_deg=(tech["angle_min"], tech["angle_max"]),
+                angle_step_deg=tech["angle_step"],
+                scale_range=(tech["scale_min"], tech["scale_max"]),
+                scale_step=tech["scale_step"],
+                spread_radius=tech["spread_radius"],
+                min_feature_spacing=tech.get("min_feature_spacing", 3),
+                pyramid_levels=tech["pyramid_levels"],
+                use_polarity=p.use_polarity,
+                use_gpu=p.use_gpu,
+            )
+            t0 = time.time(); n = m.train(roi_img); t_train = time.time() - t0
+            _check_trained(m, n)
+            _cache_put_matcher(key, m)
+        else:
+            n = len(m.variants); t_train = 0.0
+        nms = tech["nms_radius"] if tech["nms_radius"] > 0 else None
+        search_roi_t = tuple(p.search_roi) if p.search_roi else None
+        t0 = time.time()
+        matches = m.find(
+            scene, min_score=tech["min_score"], max_matches=tech["max_matches"],
+            nms_radius=nms, verify_threshold=tech["verify_threshold"],
+            # filtro_fp="off" → verify NCC davvero disabilitato
+            verify_ncc=tech.get("verify_ncc", True),
+            scale_penalty=tech.get("scale_penalty", 0.0),
+            # Halcon-mode flags
+            min_recall=p.min_recall,
+            use_soft_score=p.use_soft_score,
+            subpixel_lm=p.subpixel_lm,
+            nms_iou_threshold=p.nms_iou_threshold,
+            coarse_stride=p.coarse_stride,
+            pyramid_propagate=p.pyramid_propagate,
+            greediness=p.greediness,
+            refine_pose_joint=p.refine_pose_joint,
+            search_roi=search_roi_t,
         )
-        t0 = time.time(); n = m.train(roi_img); t_train = time.time() - t0
-        _cache_put_matcher(key, m)
-    else:
-        n = len(m.variants); t_train = 0.0
-    nms = tech["nms_radius"] if tech["nms_radius"] > 0 else None
-    search_roi_t = tuple(p.search_roi) if p.search_roi else None
-    t0 = time.time()
-    matches = m.find(
-        scene, min_score=tech["min_score"], max_matches=tech["max_matches"],
-        nms_radius=nms, verify_threshold=tech["verify_threshold"],
-        scale_penalty=tech.get("scale_penalty", 0.0),
-        # Halcon-mode flags
-        min_recall=p.min_recall,
-        use_soft_score=p.use_soft_score,
-        subpixel_lm=p.subpixel_lm,
-        nms_iou_threshold=p.nms_iou_threshold,
-        coarse_stride=p.coarse_stride,
-        pyramid_propagate=p.pyramid_propagate,
-        greediness=p.greediness,
-        refine_pose_joint=p.refine_pose_joint,
-        search_roi=search_roi_t,
-    )
-    t_find = time.time() - t0
+        t_find = time.time() - t0
 
-    tg = cv2.cvtColor(roi_img, cv2.COLOR_BGR2GRAY)
-    annotated = _draw_matches(scene, matches, tg, matcher=m)
+        tg = cv2.cvtColor(roi_img, cv2.COLOR_BGR2GRAY)
+        annotated = _draw_matches(scene, matches, tg, matcher=m)
     ann_id = _store_image(annotated)
 
     return MatchResp(
@@ -672,6 +764,7 @@ def tune(p: TuneParams):
     if model is None:
         raise HTTPException(404, "Immagine non trovata")
     x, y, w, h = p.roi
+    x, y, w, h = _clamp_roi(x, y, w, h, model.shape[1], model.shape[0])
     roi_img = model[y:y + h, x:x + w]
     t = auto_tune(roi_img)
     # Esponi parametri tecnici + meta diagnostica (_self_score, _validation,
@@ -763,26 +856,23 @@ def preview_edges(p: EdgePreviewParams):
         b = int(fb[i])
         col = bin_colors[b % len(bin_colors)]
         cv2.circle(out, (int(fx[i]), int(fy[i])), 2, col, -1, cv2.LINE_AA)
-    # UCS sul baricentro feature (richiesta utente): assi X rosso, Y verde
-    bary_cx = bary_cy = None
-    if len(fx) > 0:
-        bary_cx = float(np.mean(fx))
-        bary_cy = float(np.mean(fy))
-        bx, by = int(round(bary_cx)), int(round(bary_cy))
-        axis_len = max(20, int(0.15 * max(out.shape[:2])))
-        # X axis (rosso, verso destra)
-        cv2.arrowedLine(out, (bx, by), (bx + axis_len, by),
-                        (0, 0, 255), 2, cv2.LINE_AA, tipLength=0.2)
-        cv2.putText(out, "X", (bx + axis_len + 4, by + 5),
-                    cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 255), 1, cv2.LINE_AA)
-        # Y axis (verde, verso il basso = convenzione image y-down)
-        cv2.arrowedLine(out, (bx, by), (bx, by + axis_len),
-                        (0, 255, 0), 2, cv2.LINE_AA, tipLength=0.2)
-        cv2.putText(out, "Y", (bx + 4, by + axis_len + 12),
-                    cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 1, cv2.LINE_AA)
-        # Origine: cerchio bianco con bordo nero
-        cv2.circle(out, (bx, by), 4, (0, 0, 0), -1, cv2.LINE_AA)
-        cv2.circle(out, (bx, by), 3, (255, 255, 255), -1, cv2.LINE_AA)
+    # UCS sul CENTRO ROI (coerente con _draw_matches che usa centro pose).
+    # In questo modo l'UCS visualizzato nel modello = UCS del match (modulo
+    # rotazione/traslazione data dalla pose del pezzo trovato).
+    rh, rw = roi_img.shape[:2]
+    bx, by = (rw - 1) // 2, (rh - 1) // 2
+    axis_len = max(20, int(0.15 * max(rw, rh)))
+    cv2.arrowedLine(out, (bx, by), (bx + axis_len, by),
+                    (0, 0, 255), 2, cv2.LINE_AA, tipLength=0.2)
+    cv2.putText(out, "X", (bx + axis_len + 4, by + 5),
+                cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 255), 1, cv2.LINE_AA)
+    cv2.arrowedLine(out, (bx, by), (bx, by + axis_len),
+                    (0, 255, 0), 2, cv2.LINE_AA, tipLength=0.2)
+    cv2.putText(out, "Y", (bx + 4, by + axis_len + 12),
+                cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 1, cv2.LINE_AA)
+    cv2.circle(out, (bx, by), 4, (0, 0, 0), -1, cv2.LINE_AA)
+    cv2.circle(out, (bx, by), 3, (255, 255, 255), -1, cv2.LINE_AA)
+    bary_cx, bary_cy = float(bx), float(by)
     img_id = _store_image(out)
     n_edge_strong = int((mag >= m.strong_grad).sum())
     n_edge_total = int(edge_mask.sum() / 255)
@@ -808,6 +898,7 @@ def save_recipe(p: SaveRecipeParams):
     if model is None:
         raise HTTPException(404, "Modello non trovato")
     x, y, w, h = p.roi
+    x, y, w, h = _clamp_roi(x, y, w, h, model.shape[1], model.shape[0])
     roi_img = model[y:y + h, x:x + w]
     sp = SimpleMatchParams(
         model_id=p.model_id, scene_id=p.scene_id or p.model_id, roi=p.roi,
@@ -832,7 +923,10 @@ def save_recipe(p: SaveRecipeParams):
         use_polarity=p.use_polarity,
         use_gpu=p.use_gpu,
     )
-    m.train(roi_img)
+    # Lock globale: serializza il training pesante col matching in corso
+    with _MATCHER_LOCK:
+        n_var = m.train(roi_img)
+    _check_trained(m, n_var)
     safe_name = "".join(c for c in p.name if c.isalnum() or c in "._-")
     if not safe_name:
         raise HTTPException(400, "Nome ricetta non valido")
@@ -858,6 +952,14 @@ _RECIPE_MATCHERS: OrderedDict = OrderedDict()
 _RECIPE_MATCHERS_SIZE = 4
 
 
+def _recipe_matchers_put(name: str, matcher: LineShapeMatcher) -> None:
+    """Inserisce in _RECIPE_MATCHERS con eviction LRU (cap _RECIPE_MATCHERS_SIZE)."""
+    _RECIPE_MATCHERS[name] = matcher
+    _RECIPE_MATCHERS.move_to_end(name)
+    while len(_RECIPE_MATCHERS) > _RECIPE_MATCHERS_SIZE:
+        _RECIPE_MATCHERS.popitem(last=False)
+
+
 @app.post("/recipes/{name}/load")
 def load_recipe(name: str):
     """Carica ricetta .npz e popola cache matcher in memoria.
@@ -872,10 +974,8 @@ def load_recipe(name: str):
     if not path.is_file():
         raise HTTPException(404, f"Ricetta non trovata: {safe_name}")
     m = LineShapeMatcher.load_model(str(path))
-    _RECIPE_MATCHERS[safe_name] = m
-    _RECIPE_MATCHERS.move_to_end(safe_name)
-    while len(_RECIPE_MATCHERS) > _RECIPE_MATCHERS_SIZE:
-        _RECIPE_MATCHERS.popitem(last=False)
+    with _MATCHER_LOCK:
+        _recipe_matchers_put(safe_name, m)
     return {
         "name": safe_name,
         "n_variants": len(m.variants),
@@ -899,7 +999,9 @@ class RecipeMatchParams(BaseModel):
     greediness: float = 0.0
     refine_pose_joint: bool = False
     search_roi: list[int] | None = None
-    verify_threshold: float = 0.5
+    # Allineato a MatchParams.verify_threshold (0.4): valori divergenti
+    # davano risultati diversi tra /match e /match_recipe a parità di scena.
+    verify_threshold: float = 0.4
     scale_penalty: float = 0.0
 
 
@@ -907,37 +1009,44 @@ class RecipeMatchParams(BaseModel):
 def match_recipe(p: RecipeMatchParams):
     """Match con ricetta pre-trained: zero training, solo find."""
     safe_name = p.recipe if p.recipe.endswith(".npz") else f"{p.recipe}.npz"
-    m = _RECIPE_MATCHERS.get(safe_name)
-    if m is None:
-        # Auto-load on demand
-        path = RECIPES_DIR / safe_name
-        if not path.is_file():
-            raise HTTPException(404, f"Ricetta non trovata: {safe_name}")
-        m = LineShapeMatcher.load_model(str(path))
-        _RECIPE_MATCHERS[safe_name] = m
     scene = _load_image(p.scene_id)
     if scene is None:
         raise HTTPException(404, "Scena non trovata")
     search_roi_t = tuple(p.search_roi) if p.search_roi else None
-    t0 = time.time()
-    matches = m.find(
-        scene,
-        min_score=p.min_score, max_matches=p.max_matches,
-        verify_threshold=p.verify_threshold,
-        scale_penalty=p.scale_penalty,
-        min_recall=p.min_recall,
-        use_soft_score=p.use_soft_score,
-        subpixel_lm=p.subpixel_lm,
-        nms_iou_threshold=p.nms_iou_threshold,
-        coarse_stride=p.coarse_stride,
-        pyramid_propagate=p.pyramid_propagate,
-        greediness=p.greediness,
-        refine_pose_joint=p.refine_pose_joint,
-        search_roi=search_roi_t,
-    )
-    t_find = time.time() - t0
-    tg = m.template_gray if m.template_gray is not None else np.zeros((1, 1), np.uint8)
-    annotated = _draw_matches(scene, matches, tg, matcher=m)
+    # Lock globale: matcher condivisi tra thread del pool FastAPI
+    with _MATCHER_LOCK:
+        m = _RECIPE_MATCHERS.get(safe_name)
+        if m is not None:
+            _RECIPE_MATCHERS.move_to_end(safe_name)  # LRU touch
+        else:
+            # Auto-load on demand: stessa eviction LRU di load_recipe
+            # (senza cap la cache cresceva senza limite)
+            path = RECIPES_DIR / safe_name
+            if not path.is_file():
+                raise HTTPException(404, f"Ricetta non trovata: {safe_name}")
+            m = LineShapeMatcher.load_model(str(path))
+            _recipe_matchers_put(safe_name, m)
+        t0 = time.time()
+        matches = m.find(
+            scene,
+            min_score=p.min_score, max_matches=p.max_matches,
+            verify_threshold=p.verify_threshold,
+            # Soglia 0 = filtro FP disattivato: skippa proprio il calcolo NCC
+            verify_ncc=p.verify_threshold > 0.0,
+            scale_penalty=p.scale_penalty,
+            min_recall=p.min_recall,
+            use_soft_score=p.use_soft_score,
+            subpixel_lm=p.subpixel_lm,
+            nms_iou_threshold=p.nms_iou_threshold,
+            coarse_stride=p.coarse_stride,
+            pyramid_propagate=p.pyramid_propagate,
+            greediness=p.greediness,
+            refine_pose_joint=p.refine_pose_joint,
+            search_roi=search_roi_t,
+        )
+        t_find = time.time() - t0
+        tg = m.template_gray if m.template_gray is not None else np.zeros((1, 1), np.uint8)
+        annotated = _draw_matches(scene, matches, tg, matcher=m)
     ann_id = _store_image(annotated)
     return MatchResp(
         matches=[MatchResult(

pm2d/web/static/index.html

@@ -102,8 +102,8 @@
     <div class="field">
       <label>Simmetria</label>
       <select id="p-simmetria">
+        <option value="nessuna" selected>Nessuna (0..360°)</option>
         <option value="invariante">Invariante (cerchi — no rotazione)</option>
-        <option value="nessuna">Nessuna (0..360°)</option>
         <option value="bilaterale">Bilaterale (speculare 180°)</option>
         <option value="rot_3">Rotazionale 3× (120°)</option>
         <option value="rot_4">Rotazionale 4× (90°)</option>

pyproject.toml

@@ -14,6 +14,7 @@ dependencies = [
 
 [project.scripts]
 pm2d-eval = "pm2d.eval:main"
+pm2d-bench = "pm2d.bench:main"
 
 [dependency-groups]
 dev = [