Compare commits

..

1 Commits

Author SHA1 Message Date
Adriano 746d1668c6 feat: NCC verify lazy con skip per shape-score alto
ncc_skip_above (default 0.85): se lo score shape e gia molto alto,
salta la verifica NCC (costosa: warp + corr per ogni match). I match
borderline 0.6-0.85 vengono comunque verificati.

Comportamento Halcon-style: NCC come tie-breaker per casi ambigui,
non come gate generalizzato. Su scene con molti match netti riduce
sensibilmente il costo della fase post-NMS.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-04 15:28:24 +02:00
+14 -30
View File
@@ -239,8 +239,6 @@ class LineShapeMatcher:
self._train_mask = mask_full.copy() self._train_mask = mask_full.copy()
self.variants.clear() self.variants.clear()
# Invalida cache feature di refine: il template e cambiato.
self._refine_feat_cache = {}
for s in self._scale_list(): for s in self._scale_list():
sw = max(16, int(round(w * s))) sw = max(16, int(round(w * s)))
sh = max(16, int(round(h * s))) sh = max(16, int(round(h * s)))
@@ -435,36 +433,17 @@ class LineShapeMatcher:
H, W = spread0.shape H, W = spread0.shape
margin = 3 margin = 3
# Cache template features per angolo (chiave: int(round(ang*20)) =
# bucket di 0.05°). Golden-search ricontratto puo richiedere lo
# stesso bucket piu volte; evita re-warp+gradient+extract (costoso).
# Cache a livello matcher per riusare tra chiamate find() su scene
# diverse: la rotazione del template non dipende dalla scena.
if not hasattr(self, '_refine_feat_cache'):
self._refine_feat_cache = {}
feat_cache = self._refine_feat_cache
cache_scale_key = round(scale * 1000)
def _score_at_angle(off: float) -> tuple[float, float, float]: def _score_at_angle(off: float) -> tuple[float, float, float]:
"""Ritorna (score, best_cx, best_cy) per angolo = angle_deg + off.""" """Ritorna (score, best_cx, best_cy) per angolo = angle_deg + off."""
ang = angle_deg + off ang = angle_deg + off
ck = (round(ang * 20), cache_scale_key) M = cv2.getRotationMatrix2D(center, ang, 1.0)
cached = feat_cache.get(ck) gray_r = cv2.warpAffine(gray_p, M, (diag, diag),
if cached is not None: flags=cv2.INTER_LINEAR,
fx, fy, fb = cached borderMode=cv2.BORDER_REPLICATE)
else: mask_r = cv2.warpAffine(mask_p, M, (diag, diag),
M = cv2.getRotationMatrix2D(center, ang, 1.0) flags=cv2.INTER_NEAREST, borderValue=0)
gray_r = cv2.warpAffine(gray_p, M, (diag, diag), mag, bins = self._gradient(gray_r)
flags=cv2.INTER_LINEAR, fx, fy, fb = self._extract_features(mag, bins, mask_r)
borderMode=cv2.BORDER_REPLICATE)
mask_r = cv2.warpAffine(mask_p, M, (diag, diag),
flags=cv2.INTER_NEAREST, borderValue=0)
mag, bins = self._gradient(gray_r)
fx, fy, fb = self._extract_features(mag, bins, mask_r)
# LRU semplice: limita cache a ~256 angoli (8 angoli * 32 candidati)
if len(feat_cache) > 256:
feat_cache.pop(next(iter(feat_cache)))
feat_cache[ck] = (fx, fy, fb)
if len(fx) < 8: if len(fx) < 8:
return (0.0, cx, cy) return (0.0, cx, cy)
dx = (fx - center[0]).astype(np.int32) dx = (fx - center[0]).astype(np.int32)
@@ -593,6 +572,7 @@ class LineShapeMatcher:
subpixel: bool = True, subpixel: bool = True,
verify_ncc: bool = True, verify_ncc: bool = True,
verify_threshold: float = 0.4, verify_threshold: float = 0.4,
ncc_skip_above: float = 0.85,
coarse_angle_factor: int = 2, coarse_angle_factor: int = 2,
scale_penalty: float = 0.0, scale_penalty: float = 0.0,
) -> list[Match]: ) -> list[Match]:
@@ -826,7 +806,11 @@ class LineShapeMatcher:
search_radius=self.angle_step_deg / 2.0, search_radius=self.angle_step_deg / 2.0,
original_score=score, original_score=score,
) )
if verify_ncc: # NCC verify lazy (Halcon-style): skip se shape-score gia molto
# alto (probabilita falso positivo trascurabile). NCC e l'op
# piu costosa per match (warp + corr), quindi vale la pena
# saltarlo quando il gradiente shape e gia conclusivo.
if verify_ncc and float(score_f) < ncc_skip_above:
ncc = self._verify_ncc(gray0, cx_f, cy_f, ang_f, var.scale) ncc = self._verify_ncc(gray0, cx_f, cy_f, ang_f, var.scale)
if ncc < verify_threshold: if ncc < verify_threshold:
continue continue