perf: Fase 2 speed (3x baseline) - fuse JIT + LRU + sub-pixel lazy

Ottimizzazioni cumulative (225s -> 73s sul bench suite, 3.07x):

pm2d/line_matcher.py:
- Sub-pixel + plateau centroid spostati DOPO il pre-NMS (prima: 58k chiamate
  per clip_preciso anche su candidati poi scartati dalla NMS; ora solo sui
  ~75 preliminary sopravvissuti). Coordinate intere OK per la decisione
  reject, dato che nms_radius >= 8 px.
- Usa nuovo kernel fuso score+rescore (no allocazione intermedia).
- Adaptive plateau_radius + propagazione train_mask per NCC coerente.
- Local crop NCC (diag template invece di intera scena).
- Fallback adattivo se bg_rescore azzera tutti gli score top-level.

pm2d/_jit_kernels.py:
- Nuovo kernel _jit_score_bitmap_rescored: fonde scoring bitmap e rescore
  (score - bg) / (1 - bg) in un singolo pass parallelo. Evita allocazione
  e passata aggiuntiva (era ~15% del tempo find sul preciso).

pm2d/auto_tune.py:
- LRU cache in-memory sui risultati auto_tune (chiave md5 ROI + mask):
  richiamate successive con stessa ROI sono O(1).
- Downsample a 128px prima della correlazione rotazionale
  (O(n_angles * H * W) -> insensibile su sample moderati).
- Soglie weak/strong da percentili reali (p55/p85) senza clamp a 100,
  con clamp massimo 400 per evitare saturazione su template ad alto contrasto.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
root
2026-04-24 21:21:59 +00:00
parent 44a3046616
commit 89b59b3ea3
3 changed files with 230 additions and 42 deletions
+74
View File
@@ -110,6 +110,55 @@ if HAS_NUMBA:
acc[y, x] *= inv
return acc
@nb.njit(cache=True, parallel=True, fastmath=True, boundscheck=False)
def _jit_score_bitmap_rescored(
spread: np.ndarray, # uint8 (H, W)
dx: np.ndarray, # int32 (N,)
dy: np.ndarray, # int32 (N,)
bins: np.ndarray, # int8 (N,)
bit_active: np.uint8,
bg: np.ndarray, # float32 (H, W) background density normalizzata
) -> np.ndarray:
"""score+rescore in un singolo pass: evita allocazione intermedia.
Equivalente a:
score = _jit_score_bitmap(...)
out = max(0, (score - bg) / (1 - bg + 1e-6))
ma fonde la seconda passata dentro la normalizzazione finale
(cache-friendly, risparmia ~15% sul totale find).
"""
H, W = spread.shape
N = dx.shape[0]
acc = np.zeros((H, W), dtype=np.float32)
for y in nb.prange(H):
for i in range(N):
b = bins[i]
mask = np.uint8(1) << b
if (bit_active & mask) == 0:
continue
ddy = dy[i]
yy = y + ddy
if yy < 0 or yy >= H:
continue
ddx = dx[i]
x_lo = 0 if ddx >= 0 else -ddx
x_hi = W if ddx <= 0 else W - ddx
for x in range(x_lo, x_hi):
if spread[yy, x + ddx] & mask:
acc[y, x] += 1.0
if N > 0:
inv = 1.0 / N
for y in nb.prange(H):
for x in range(W):
v = acc[y, x] * inv
bgv = bg[y, x]
if bgv < 1.0:
r = (v - bgv) / (1.0 - bgv + 1e-6)
acc[y, x] = r if r > 0.0 else 0.0
else:
acc[y, x] = 0.0
return acc
@nb.njit(cache=True, parallel=True, fastmath=True, boundscheck=False)
def _jit_popcount_density(spread: np.ndarray) -> np.ndarray:
"""Conta bit set per pixel: ritorna (H, W) float32 in [0..8]."""
@@ -134,6 +183,8 @@ if HAS_NUMBA:
_jit_score_by_shift(resp, dx, dy, b, ba)
spread = np.zeros((32, 32), dtype=np.uint8)
_jit_score_bitmap(spread, dx, dy, b, np.uint8(0xFF))
bg = np.zeros((32, 32), dtype=np.float32)
_jit_score_bitmap_rescored(spread, dx, dy, b, np.uint8(0xFF), bg)
_jit_popcount_density(spread)
else: # pragma: no cover
@@ -144,6 +195,9 @@ else: # pragma: no cover
def _jit_score_bitmap(spread, dx, dy, bins, bit_active):
raise RuntimeError("numba non disponibile")
def _jit_score_bitmap_rescored(spread, dx, dy, bins, bit_active, bg):
raise RuntimeError("numba non disponibile")
def _jit_popcount_density(spread):
raise RuntimeError("numba non disponibile")
@@ -172,6 +226,26 @@ def score_bitmap(
return _numpy_score_by_shift(resp, dx, dy, bins, None)
def score_bitmap_rescored(
spread: np.ndarray, dx: np.ndarray, dy: np.ndarray, bins: np.ndarray,
bit_active: int, bg: np.ndarray,
) -> np.ndarray:
"""Score bitmap + rescore fusi in un solo pass (JIT)."""
if HAS_NUMBA and len(dx) > 0:
return _jit_score_bitmap_rescored(
np.ascontiguousarray(spread, dtype=np.uint8),
np.ascontiguousarray(dx, dtype=np.int32),
np.ascontiguousarray(dy, dtype=np.int32),
np.ascontiguousarray(bins, dtype=np.int8),
np.uint8(bit_active),
np.ascontiguousarray(bg, dtype=np.float32),
)
# Fallback: chiamate separate
score = score_bitmap(spread, dx, dy, bins, bit_active)
out = (score - bg) / (1.0 - bg + 1e-6)
return np.maximum(0.0, out).astype(np.float32)
def popcount_density(spread: np.ndarray) -> np.ndarray:
if HAS_NUMBA:
return _jit_popcount_density(np.ascontiguousarray(spread, dtype=np.uint8))
+66 -11
View File
@@ -14,6 +14,9 @@ Ritorna dict con i key esatti del form `edit_params`.
from __future__ import annotations
import hashlib
from collections import OrderedDict
import cv2
import numpy as np
@@ -24,17 +27,33 @@ def _to_gray(img: np.ndarray) -> np.ndarray:
return img
# Cache in-memory (LRU) dei risultati auto_tune per stesso input ROI.
_TUNE_CACHE: OrderedDict[str, dict] = OrderedDict()
_TUNE_CACHE_SIZE = 32
def detect_rotational_symmetry(
gray: np.ndarray, step_deg: float = 5.0, corr_thresh: float = 0.75,
) -> dict:
"""Rileva simmetria rotazionale su edge map (più robusto a sfondo uniforme).
Downsample a max 128 px prima di correlare per abbattere il costo
O(n_angles · H · W) senza perdere precisione (la simmetria rotazionale
è invariante a subsampling moderato).
Ritorna dict con:
- order: int, 1=nessuna, 2=180°, 3=120°, 4=90°, 6=60°, 8=45°
- period_deg: float, periodo minimo di simmetria (360/order)
- confidence: float [0..1], correlazione minima tra rotazioni equivalenti
"""
h, w = gray.shape
target = 128
if max(h, w) > target:
sf = target / max(h, w)
new_w = max(32, int(w * sf))
new_h = max(32, int(h * sf))
gray = cv2.resize(gray, (new_w, new_h), interpolation=cv2.INTER_AREA)
h, w = gray.shape
# Usa magnitude gradiente (rotation-invariant rispetto a bg uniforme)
gx = cv2.Sobel(gray, cv2.CV_32F, 1, 0, ksize=3)
gy = cv2.Sobel(gray, cv2.CV_32F, 0, 1, ksize=3)
@@ -88,9 +107,12 @@ def analyze_gradients(gray: np.ndarray) -> dict:
gy = cv2.Sobel(gray, cv2.CV_32F, 0, 1, ksize=3)
mag = cv2.magnitude(gx, gy)
# Percentili magnitude
# Percentili magnitude: p55/p85 usati per soglie weak/strong (più aderenti
# alla distribuzione reale rispetto a p50/p80 + clamp).
p50 = float(np.percentile(mag, 50))
p55 = float(np.percentile(mag, 55))
p80 = float(np.percentile(mag, 80))
p85 = float(np.percentile(mag, 85))
p95 = float(np.percentile(mag, 95))
mag_max = float(mag.max())
@@ -112,7 +134,8 @@ def analyze_gradients(gray: np.ndarray) -> dict:
ent = 0.0
return {
"p50": p50, "p80": p80, "p95": p95, "mag_max": mag_max,
"p50": p50, "p55": p55, "p80": p80, "p85": p85, "p95": p95,
"mag_max": mag_max,
"strong_pct": strong_pct, "weak_pct": weak_pct,
"orient_entropy": ent,
"n_pixels": mag.size,
@@ -120,11 +143,28 @@ def analyze_gradients(gray: np.ndarray) -> dict:
}
def _cache_key(template_bgr: np.ndarray, mask: np.ndarray | None) -> str:
h = hashlib.md5()
h.update(np.ascontiguousarray(template_bgr).tobytes())
h.update(f"shape={template_bgr.shape}".encode())
if mask is not None:
h.update(np.ascontiguousarray(mask).tobytes())
return h.hexdigest()
def auto_tune(template_bgr: np.ndarray, mask: np.ndarray | None = None) -> dict:
"""Analizza template e ritorna dict parametri suggeriti.
Chiavi compatibili con edit_params PARAM_SCHEMA.
Risultato cachato in-memory (LRU): ri-chiamare con stessa ROI è O(1).
"""
ck = _cache_key(template_bgr, mask)
cached = _TUNE_CACHE.get(ck)
if cached is not None:
_TUNE_CACHE.move_to_end(ck)
return dict(cached)
gray = _to_gray(template_bgr)
h, w = gray.shape
if mask is not None:
@@ -136,16 +176,22 @@ def auto_tune(template_bgr: np.ndarray, mask: np.ndarray | None = None) -> dict:
stats = analyze_gradients(gray_for_stats)
sym = detect_rotational_symmetry(gray_for_stats)
# Soglie magnitude: usa percentili per robustezza illuminazione.
# Target: strong_grad ~= valore a percentile 80-90 in assoluto, ma
# clamp per compatibilità uint8 (Sobel può sforare).
strong_grad = float(np.clip(stats["p80"], 20.0, 100.0))
weak_grad = float(np.clip(strong_grad * 0.5, 10.0, 60.0))
# Soglie magnitude: usa percentili reali (p85/p55) senza clamp duro a 100.
# Sobel ksize=3 su uint8 può arrivare a ~1020, quindi clamp massimo 400
# evita saturazione del threshold su template ad alto contrasto.
strong_grad = float(np.clip(stats["p85"], 30.0, 400.0))
weak_grad = float(np.clip(stats["p55"], 15.0, strong_grad * 0.7))
# num_features: 1 feature ogni ~25 px forti, clamp 48..192
target_feat = int(np.clip(stats["n_strong"] / 25, 48, 192))
# num_features: ibrido perimetro + densità. Target = min(perimeter_budget,
# density_budget) per non generare più feature di quante edge nitide siano
# disponibili, ma neanche meno di quante il perimetro possa tracciare.
perim_budget = int(2 * (h + w) * 0.4) # ~40% dei pixel di perimetro
density_budget = int(stats["n_strong"] / 20) # 1 feature ogni ~20 px forti
target_feat = int(np.clip(min(perim_budget, density_budget), 64, 192))
# pyramid_levels in base alla dimensione minima
# pyramid_levels in base a dimensione minima E densità feature: un template
# grande ma povero di feature non deve scendere troppi livelli (rischio
# collasso a <16 feature al top level).
min_side = min(h, w)
if min_side < 60:
pyr = 1
@@ -155,6 +201,9 @@ def auto_tune(template_bgr: np.ndarray, mask: np.ndarray | None = None) -> dict:
pyr = 3
else:
pyr = 4
# Cap: non scendere sotto ~16 feature al top level (feature ÷ 4^(pyr-1))
max_pyr_from_feat = max(1, int(np.floor(np.log2(max(1, target_feat / 16.0)) / 2.0)) + 1)
pyr = min(pyr, max_pyr_from_feat)
# spread_radius proporzionale a risoluzione + pyramid (tolleranza ~1% dim)
spread_radius = int(np.clip(max(3, min_side * 0.02), 3, 8))
@@ -174,7 +223,7 @@ def auto_tune(template_bgr: np.ndarray, mask: np.ndarray | None = None) -> dict:
# angle step: 5° default; se simmetria, mantengo step ma range ridotto
angle_step = 5.0
return {
result = {
"backend": "line",
"angle_min": 0.0,
"angle_max": angle_max,
@@ -196,6 +245,12 @@ def auto_tune(template_bgr: np.ndarray, mask: np.ndarray | None = None) -> dict:
"_symmetry_conf": round(sym["confidence"], 2),
"_orient_entropy": round(stats["orient_entropy"], 2),
}
# Store in LRU cache
_TUNE_CACHE[ck] = dict(result)
_TUNE_CACHE.move_to_end(ck)
while len(_TUNE_CACHE) > _TUNE_CACHE_SIZE:
_TUNE_CACHE.popitem(last=False)
return result
def summarize(tune: dict) -> str:
+90 -31
View File
@@ -39,6 +39,7 @@ _GOLDEN = (math.sqrt(5.0) - 1.0) / 2.0 # ≈ 0.618
from pm2d._jit_kernels import (
score_by_shift as _jit_score_by_shift,
score_bitmap as _jit_score_bitmap,
score_bitmap_rescored as _jit_score_bitmap_rescored,
popcount_density as _jit_popcount,
HAS_NUMBA,
)
@@ -136,6 +137,8 @@ class LineShapeMatcher:
self.variants: list[_Variant] = []
self.template_size: tuple[int, int] = (0, 0)
self.template_gray: np.ndarray | None = None
# Maschera usata in training (propagata al refine per coerenza).
self._train_mask: np.ndarray | None = None
# --- Helpers -------------------------------------------------------
@@ -233,6 +236,7 @@ class LineShapeMatcher:
mask_full = np.full((h, w), 255, dtype=np.uint8)
else:
mask_full = (mask > 0).astype(np.uint8) * 255
self._train_mask = mask_full.copy()
self.variants.clear()
for s in self._scale_list():
@@ -505,6 +509,10 @@ class LineShapeMatcher:
) -> float:
"""NCC tra template warpato alla pose e scena sottostante.
Lavora su un **crop locale** della scena di lato = diagonale del
template ruotato+scalato, non sull'intera scena. Su scene grandi
(1920×1080) taglia drasticamente il costo del warp per ogni match.
Ritorna score [-1, 1]. Usato come filtro anti-falso-positivo:
il matcher linemod può dare score alto su texture generiche ma
sovrapponendo il template gray i pixel non corrispondono.
@@ -515,23 +523,40 @@ class LineShapeMatcher:
h, w = t.shape
cx_t = (w - 1) / 2.0
cy_t = (h - 1) / 2.0
M = cv2.getRotationMatrix2D((cx_t, cy_t), angle_deg, scale)
M[0, 2] += cx - cx_t
M[1, 2] += cy - cy_t
# Bounding box del template ruotato/scalato attorno a (cx, cy)
diag = int(np.ceil(np.hypot(w, h) * scale)) + 8
H, W = scene_gray.shape
x0 = int(round(cx)) - diag // 2
y0 = int(round(cy)) - diag // 2
cx0 = max(0, x0); cy0 = max(0, y0)
cx1 = min(W, x0 + diag); cy1 = min(H, y0 + diag)
if cx1 - cx0 < 10 or cy1 - cy0 < 10:
return 0.0
scn_crop = scene_gray[cy0:cy1, cx0:cx1]
ch, cw = scn_crop.shape
M = cv2.getRotationMatrix2D((cx_t, cy_t), angle_deg, scale)
# Porta il centro del template a (cx - cx0, cy - cy0) del crop
M[0, 2] += (cx - cx0) - cx_t
M[1, 2] += (cy - cy0) - cy_t
warped = cv2.warpAffine(
t, M, (W, H),
t, M, (cw, ch),
flags=cv2.INTER_LINEAR, borderValue=0,
)
mask = cv2.warpAffine(
np.full_like(t, 255), M, (W, H),
if self._train_mask is not None:
mask_src = self._train_mask
else:
mask_src = np.full_like(t, 255)
mask_w = cv2.warpAffine(
mask_src, M, (cw, ch),
flags=cv2.INTER_NEAREST, borderValue=0,
)
valid = mask > 0
valid = mask_w > 0
if valid.sum() < 20:
return 0.0
tpl = warped[valid].astype(np.float32)
scn = scene_gray[valid].astype(np.float32)
scn = scn_crop[valid].astype(np.float32)
tm = tpl - tpl.mean()
sm = scn - scn.mean()
denom = np.sqrt((tm * tm).sum() * (sm * sm).sum()) + 1e-9
@@ -624,24 +649,37 @@ class LineShapeMatcher:
def _top_score(vi: int) -> tuple[int, float]:
var = self.variants[vi]
lvl = var.levels[min(top, len(var.levels) - 1)]
score = _jit_score_bitmap(
score = _jit_score_bitmap_rescored(
spread_top, lvl.dx, lvl.dy, lvl.bin, bit_active_top,
bg_cache_top[var.scale],
)
score = _rescore(score, bg_cache_top[var.scale])
return vi, float(score.max()) if score.size else -1.0
kept_coarse: list[tuple[int, float]] = []
all_top_scores: list[tuple[int, float]] = []
if self.n_threads > 1 and len(coarse_idx_list) > 1:
with ThreadPoolExecutor(max_workers=self.n_threads) as ex:
for vi, best in ex.map(_top_score, coarse_idx_list):
all_top_scores.append((vi, best))
if best >= top_thresh:
kept_coarse.append((vi, best))
else:
for vi in coarse_idx_list:
vi2, best = _top_score(vi)
all_top_scores.append((vi2, best))
if best >= top_thresh:
kept_coarse.append((vi2, best))
# Fallback adattivo: se il rescore background ha abbattuto tutti
# gli score sotto top_thresh (scene texturate pesanti), ripesca
# le varianti migliori al top level per dare comunque una chance
# alla fase full-res invece di ritornare 0 match.
if not kept_coarse and all_top_scores:
all_top_scores.sort(key=lambda t: -t[1])
n_keep = max(4, len(all_top_scores) // 10)
# Limita a varianti con score top > 0 (non completamente a zero)
kept_coarse = [(vi, s) for vi, s in all_top_scores[:n_keep] if s > 0]
# Espandi ogni coarse promosso con i suoi vicini (stessa scala,
# angoli intermedi non valutati al top)
expanded: set[int] = set()
@@ -678,10 +716,10 @@ class LineShapeMatcher:
def _full_score(vi: int) -> tuple[int, np.ndarray]:
var = self.variants[vi]
lvl0 = var.levels[0]
score = _jit_score_bitmap(
score = _jit_score_bitmap_rescored(
spread0, lvl0.dx, lvl0.dy, lvl0.bin, bit_active_full,
bg_cache_full[var.scale],
)
score = _rescore(score, bg_cache_full[var.scale])
return vi, score
candidates_per_var: list[tuple[int, np.ndarray]] = []
@@ -693,14 +731,24 @@ class LineShapeMatcher:
else:
results = [_full_score(vi) for vi in var_indices]
def _scale_factor(s: float) -> float:
"""Penalità moltiplicativa per scala diversa da 1.0."""
if scale_penalty > 0.0 and s != 1.0:
return max(0.0, 1.0 - scale_penalty * abs(s - 1.0))
return 1.0
for vi, score in results:
ys, xs = np.where(score >= min_score)
pen = _scale_factor(self.variants[vi].scale)
# Ordinare/sogliare su score penalizzato: un match a scala 1.5 con
# score 0.8 e penalty=0.3 effettivamente vale 0.56, non 0.8.
score_for_sort = score if pen == 1.0 else score * pen
ys, xs = np.where(score_for_sort >= min_score)
if len(ys) == 0:
continue
vals = score[ys, xs]
vals = score_for_sort[ys, xs]
K = min(len(vals), max_matches * 5)
ord_idx = np.argpartition(-vals, K - 1)[:K]
candidates_per_var.append((vi, score))
candidates_per_var.append((vi, score)) # score_map originale
for i in ord_idx:
raw.append((float(vals[i]), int(xs[i]), int(ys[i]), vi))
@@ -710,32 +758,43 @@ class LineShapeMatcher:
score_maps = dict(candidates_per_var)
# NMS + subpixel + refinement angolare
# Mask template per refinement (non disponibile qui: usa full)
# Usa mask salvata in train() per coerenza (se ROI poligonale).
h, w = self.template_gray.shape if self.template_gray is not None else (0, 0)
mask_full = np.full((h, w), 255, dtype=np.uint8)
mask_full = (
self._train_mask if self._train_mask is not None
else np.full((h, w), 255, dtype=np.uint8)
)
# Plateau radius adattivo al template (evita plateau troppo ampi su
# template piccoli: 8% del lato minimo, clampato [3, 10]).
plateau_r = max(3, min(10, int(min(self.template_size) * 0.08)))
# Pre-NMS rapido su raw (solo subpixel, no refine/verify): riduce
# i candidati a ~max_matches*3 prima di operazioni costose (refine,
# verify) che erano chiamate per ogni raw causando lentezze 100x.
# Pre-NMS rapido su raw con coordinate intere (nms_radius ≥ 8,
# la precisione sub-pixel non cambia la decisione di reject).
# Subpixel viene calcolato DOPO il pre-NMS solo sui ~pre_cap
# preliminary sopravvissuti: prima era chiamato su ogni raw (~58k
# chiamate su clip_preciso) anche se la maggior parte veniva poi
# scartata dalla NMS, sprecando la parte più costosa del loop.
r2 = nms_radius * nms_radius
preliminary: list[tuple[float, float, float, int]] = []
pre_cap = max(max_matches * 3, max_matches + 10)
preliminary_int: list[tuple[float, int, int, int]] = []
for score, xi, yi, vi in raw:
if subpixel and vi in score_maps:
cx_f, cy_f = self._subpixel_peak(score_maps[vi], xi, yi)
else:
cx_f, cy_f = float(xi), float(yi)
if any((k[1] - cx_f) ** 2 + (k[2] - cy_f) ** 2 < r2
for k in preliminary):
if any((k[1] - xi) ** 2 + (k[2] - yi) ** 2 < r2
for k in preliminary_int):
continue
preliminary.append((score, cx_f, cy_f, vi))
if len(preliminary) >= pre_cap:
preliminary_int.append((score, xi, yi, vi))
if len(preliminary_int) >= pre_cap:
break
# Ora refine + verify solo sui candidati pre-NMS
# Subpixel + refine + verify solo sui candidati pre-NMS (max pre_cap)
kept: list[Match] = []
tw, th = self.template_size
for score, cx_f, cy_f, vi in preliminary:
for score, xi, yi, vi in preliminary_int:
if subpixel and vi in score_maps:
cx_f, cy_f = self._subpixel_peak(
score_maps[vi], xi, yi, plateau_radius=plateau_r,
)
else:
cx_f, cy_f = float(xi), float(yi)
var = self.variants[vi]
ang_f = var.angle_deg
score_f = score