1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
|
"""CodeShell.kr — misc-mosaic: compact-rectangle jigsaw assembly.
Free-form frontier growth sprawls: slots with a single constraint can look
confident yet be wrong, so the blob drifts and ends up 49x69 instead of 24x24.
This version keeps the assembly a solid rectangle: it repeatedly fills every
empty cell inside the current bounding box (highest confidence first), and only
then expands the box by one row/column on whichever side matches best.
"""
import argparse
from pathlib import Path
import numpy as np
from PIL import Image
PIECES = "extracted/misc-mosaic/pieces"
OUT = "extracted/misc-mosaic/assembled3.png"
class Assembler:
def __init__(self, rows, cols, minconf, maxcost):
ps = sorted(Path(PIECES).glob("*.png"))
arrs = [np.array(Image.open(p).convert("RGB"), dtype=np.int32) for p in ps]
self.O = np.stack([np.rot90(a, k=-r) for a in arrs for r in range(4)])
m = len(self.O)
self.L = self.O[:, :, 0, :].reshape(m, -1)
self.R = self.O[:, :, -1, :].reshape(m, -1)
self.T = self.O[:, 0, :, :].reshape(m, -1)
self.B = self.O[:, -1, :, :].reshape(m, -1)
self.owner = np.arange(m) // 4
self.m = m
self.avail = np.ones(m, dtype=bool)
self.grid = {}
self.rows, self.cols = rows, cols
self.minconf = minconf
self.maxcost = maxcost
def best_for(self, r, c):
left = self.grid.get((r, c - 1))
right = self.grid.get((r, c + 1))
top = self.grid.get((r - 1, c))
bottom = self.grid.get((r + 1, c))
cost = np.zeros(self.m, dtype=np.int64)
ncon = 0
if left is not None:
cost += np.abs(self.L - self.R[left]).sum(axis=1)
ncon += 1
if right is not None:
cost += np.abs(self.R - self.L[right]).sum(axis=1)
ncon += 1
if top is not None:
cost += np.abs(self.T - self.B[top]).sum(axis=1)
ncon += 1
if bottom is not None:
cost += np.abs(self.B - self.T[bottom]).sum(axis=1)
ncon += 1
if ncon == 0:
return None
cost = cost / ncon
cost[~self.avail] = 10**15
j = int(cost.argmin())
bcost = int(cost[j])
cost[self.owner == self.owner[j]] = 10**15
second = int(cost.min())
return j, bcost, (second + 1) / (bcost + 1), ncon
def place(self, r, c, j):
self.grid[(r, c)] = j
self.avail[self.owner == self.owner[j]] = False
def fill_box(self, r0, r1, c0, c1):
placed_any = False
while True:
best = None
for r in range(r0, r1 + 1):
for c in range(c0, c1 + 1):
if (r, c) in self.grid:
continue
res = self.best_for(r, c)
if res is None:
continue
j, bcost, conf, ncon = res
if conf < self.minconf or bcost > self.maxcost:
continue
score = (ncon, conf)
if best is None or score > best[0]:
best = (score, r, c, j, bcost, conf)
if best is None:
break
_, r, c, j, bcost, conf = best
self.place(r, c, j)
placed_any = True
return placed_any
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--rows", type=int, default=24)
ap.add_argument("--cols", type=int, default=24)
ap.add_argument("--minconf", type=float, default=1.15)
ap.add_argument("--maxcost", type=float, default=1500.0)
ap.add_argument("--out", default=OUT)
ap.add_argument("--seed", type=int, default=0,
help="pick the Nth most confident seed pair")
args = ap.parse_args()
A = Assembler(args.rows, args.cols, args.minconf, args.maxcost)
print(f"pieces {len(A.owner)//4} oriented {A.m}")
seeds = []
for i in range(0, A.m, 11):
d = np.abs(A.L - A.R[i]).sum(axis=1)
d[A.owner == A.owner[i]] = 10**15
j = int(d.argmin())
bcost = int(d[j])
d[A.owner == A.owner[j]] = 10**15
conf = (int(d.min()) + 1) / (bcost + 1)
seeds.append((conf, i, j))
seeds.sort(reverse=True)
conf, a, b = seeds[min(args.seed, len(seeds) - 1)]
print(f"seed #{args.seed} {a}->{b} conf {conf:.2f}")
A.place(0, 0, a)
A.place(0, 1, b)
r0 = r1 = 0
c0, c1 = 0, 1
total = args.rows * args.cols
while len(A.grid) < total:
A.fill_box(r0, r1, c0, c1)
if len(A.grid) >= total:
break
options = []
sides = {
"top": [(r0 - 1, cc) for cc in range(c0, c1 + 1)],
"bottom": [(r1 + 1, cc) for cc in range(c0, c1 + 1)],
"left": [(rr, c0 - 1) for rr in range(r0, r1 + 1)],
"right": [(rr, c1 + 1) for rr in range(r0, r1 + 1)],
}
if r1 - r0 + 1 >= args.rows:
sides.pop("top", None)
sides.pop("bottom", None)
if c1 - c0 + 1 >= args.cols:
sides.pop("left", None)
sides.pop("right", None)
for side, cells in sides.items():
best = None
for (rr, cc) in cells:
if (rr, cc) in A.grid:
continue
res = A.best_for(rr, cc)
if res is None:
continue
j, bcost, cf, ncon = res
if cf < A.minconf or bcost > A.maxcost:
continue
if best is None or cf > best:
best = cf
if best is not None:
options.append((best, side))
if not options:
print(f"stop: no expandable side (placed {len(A.grid)})")
break
options.sort(reverse=True)
side = options[0][1]
if side == "top":
r0 -= 1
elif side == "bottom":
r1 += 1
elif side == "left":
c0 -= 1
else:
c1 += 1
print(f"expand {side}: box {r1-r0+1}x{c1-c0+1} placed {len(A.grid)} "
f"conf {options[0][0]:.2f}")
h, w = r1 - r0 + 1, c1 - c0 + 1
print(f"placed {len(A.grid)} tiles, bbox {h}x{w} "
f"(target {args.rows}x{args.cols})")
A.minconf = 0.0
A.maxcost = float("inf")
A.fill_box(r0, r1, c0, c1)
print(f"after force-fill: {len(A.grid)} tiles")
refine(A, r0, r1, c0, c1, passes=6)
cells = [(r, c) for r in range(r0, r1 + 1) for c in range(c0, c1 + 1)]
refine_swaps(A, cells, passes=6)
canvas = np.zeros((h * 32, w * 32, 3), dtype=np.uint8)
for (r, c), o in A.grid.items():
canvas[(r - r0) * 32:(r - r0 + 1) * 32, (c - c0) * 32:(c - c0 + 1) * 32] = A.O[o]
Image.fromarray(canvas).save(args.out)
print(f"wrote {args.out} ({canvas.shape[1]}x{canvas.shape[0]})")
def cell_costs_all(A, r, c):
"""Vectorised local cost of every oriented piece at slot (r,c)."""
cost = np.zeros(A.m, dtype=np.int64)
left = A.grid.get((r, c - 1))
right = A.grid.get((r, c + 1))
top = A.grid.get((r - 1, c))
bottom = A.grid.get((r + 1, c))
if left is not None:
cost += np.abs(A.L - A.R[left]).sum(axis=1)
if right is not None:
cost += np.abs(A.R - A.L[right]).sum(axis=1)
if top is not None:
cost += np.abs(A.T - A.B[top]).sum(axis=1)
if bottom is not None:
cost += np.abs(A.B - A.T[bottom]).sum(axis=1)
return cost
def local_cost_at(A, r, c, o):
"""Edge cost of piece o sitting at slot (r,c) against placed neighbours."""
tot = 0
left = A.grid.get((r, c - 1))
right = A.grid.get((r, c + 1))
top = A.grid.get((r - 1, c))
bottom = A.grid.get((r + 1, c))
if left is not None:
tot += int(np.abs(A.L[o] - A.R[left]).sum())
if right is not None:
tot += int(np.abs(A.R[o] - A.L[right]).sum())
if top is not None:
tot += int(np.abs(A.T[o] - A.B[top]).sum())
if bottom is not None:
tot += int(np.abs(A.B[o] - A.T[bottom]).sum())
return tot
def refine_swaps(A, cells, passes=4):
"""Local search allowing both substitution and swaps between two slots."""
for p in range(passes):
pos = {}
for rc, o in A.grid.items():
pos[A.owner[o]] = rc
changed = 0
for (r, c) in cells:
cur = A.grid[(r, c)]
base = local_cost_at(A, r, c, cur)
cost = cell_costs_all(A, r, c)
for o in np.argsort(cost)[:40]:
o = int(o)
if o == cur:
break
if cost[o] >= base:
break
oc = A.owner[o]
if oc not in pos:
A.grid[(r, c)] = o
pos.pop(A.owner[cur], None)
pos[oc] = (r, c)
changed += 1
break
r2, c2 = pos[oc]
if (r2, c2) == (r, c):
break
other = A.grid[(r2, c2)]
base2 = local_cost_at(A, r2, c2, other)
A.grid[(r, c)] = o
A.grid[(r2, c2)] = cur
if local_cost_at(A, r, c, o) + local_cost_at(A, r2, c2, cur) < base + base2:
pos[oc] = (r, c)
pos[A.owner[cur]] = (r2, c2)
changed += 1
break
A.grid[(r, c)] = cur
A.grid[(r2, c2)] = other
print(f"swap pass {p}: {changed} slots changed")
if changed == 0:
break
def refine(A, r0, r1, c0, c1, passes=6):
"""Local search: reassign cells to the best-fitting piece/rotation."""
cells = [(r, c) for r in range(r0, r1 + 1) for c in range(c0, c1 + 1)]
for p in range(passes):
occupied = {}
for rc, o in A.grid.items():
occupied[A.owner[o]] = rc
changed = 0
for (r, c) in cells:
cur = A.grid[(r, c)]
cost = cell_costs_all(A, r, c)
order = np.argsort(cost)
pick = None
for o in order[:60]:
o = int(o)
if o == cur:
pick = cur
break
oc = A.owner[o]
if oc in occupied and occupied[oc] != (r, c):
continue
pick = o
break
if pick is None:
continue
if pick != cur:
A.grid[(r, c)] = pick
occupied.pop(A.owner[cur], None)
occupied[A.owner[pick]] = (r, c)
changed += 1
print(f"refine pass {p}: {changed} cells changed")
if changed == 0:
break
if __name__ == "__main__":
main()
|