File size: 8,642 Bytes
954caab |
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 |
from pathlib import Path
import numpy as np
import torch
import torchvision.transforms.functional as TF
from einops import rearrange, repeat
from .jigsaw_helpers import get_jigsaw_pieces
def get_inv_perm(perm):
'''
Get the inverse permutation of a permutation. That is, the array such that
perm[perm_inv] = perm_inv[perm] = arange(len(perm))
perm (torch.tensor) :
A 1-dimensional integer array, representing a permutation. Indicates
that element i should move to index perm[i]
'''
perm_inv = torch.empty_like(perm)
perm_inv[perm] = torch.arange(len(perm))
return perm_inv
def make_inner_circle_perm(im_size=64, r=24):
'''
Makes permutations for "inner circle" view. Given size of image, and
`r`, the radius of the circle. We do this by iterating through every
pixel and figuring out where it should go.
'''
perm = [] # Permutation array
# Iterate through all positions, in order
for iy in range(im_size):
for ix in range(im_size):
# Get coordinates, with origin at (0, 0)
x = ix - im_size // 2 + 0.5
y = iy - im_size // 2 + 0.5
# Do 180 deg rotation if in circle
if x**2 + y**2 < r**2:
x = -x
y = -y
# Convert back to integer coordinates
x = int(x + im_size // 2 - 0.5)
y = int(y + im_size // 2 - 0.5)
# Append destination pixel index to permutation
perm.append(x + y * im_size)
perm = torch.tensor(perm)
return perm
def make_jigsaw_perm(size, seed=0):
'''
Returns a permutation of pixels that is a jigsaw permutation
There are 3 types of pieces: corner, edge, and inner pieces. These were
created in MS Paint. They are all identical and laid out like:
c0 e0 f0 c1
f3 i0 i1 e1
e3 i3 i2 f1
c3 f2 e2 c2
where c is "corner," i is "inner," and "e" and "f" are "edges."
"e" and "f" pieces are identical, but labeled differently such that
to move any piece to the next index you can apply a 90 deg rotation.
Pieces c0, e0, f0, and i0 are defined by pngs, and will be loaded in. All
other pieces are obtained by 90 deg rotations of these "base" pieces.
Permutations are defined by:
1. permutation of corner (c) pieces (length 4 perm list)
2. permutation of inner (i) pieces (length 4 perm list)
3. permutation of edge (e) pieces (length 4 perm list)
4. permutation of edge (f) pieces (length 4 perm list)
5. list of four swaps, indicating swaps between e and f
edge pieces along the same edge (length 4 bit list)
Note these perm indexes will just be a "rotation index" indicating
how many 90 deg rotations to apply to the base pieces. The swaps
ensure that any edge piece can go to any edge piece, and are indexed
by the indexes of the "e" and "f" pieces on the edge.
Also note, order of indexes in permutation array is raster scan order. So,
go along x's first, then y's. This means y * size + x gives us the
1-D location in the permutation array. And image arrays are in
(y,x) order.
Plan of attack for making a pixel permutation array that represents
a jigsaw permutation:
1. Iterate through all pixels (in raster scan order)
2. Figure out which puzzle piece it is in initially
3. Look at the permutations, and see where it should go
4. Additionally, see if it's an edge piece, and needs to be swapped
5. Add the new (1-D) index to the permutation array
'''
np.random.seed(seed)
# Get location of puzzle pieces
piece_dir = Path(__file__).parent / 'assets'
# Get random permutations of groups of 4, and cat
identity = np.arange(4)
perm_corner = np.random.permutation(identity)
perm_inner = np.random.permutation(identity)
perm_edge1 = np.random.permutation(identity)
perm_edge2 = np.random.permutation(identity)
edge_swaps = np.random.randint(2, size=4)
piece_perms = np.concatenate([perm_corner, perm_inner, perm_edge1, perm_edge2])
# Get all 16 jigsaw pieces (in the order above)
pieces = get_jigsaw_pieces(size)
# Make permutation array to fill
perm = []
# For each pixel, figure out where it should go
for y in range(size):
for x in range(size):
# Figure out which piece (x,y) is in:
piece_idx = pieces[:,y,x].argmax()
# Figure out how many 90 deg rotations are on the piece
rot_idx = piece_idx % 4
# The perms tells us how many 90 deg rotations to apply to
# arrive at new pixel location
dest_rot_idx = piece_perms[piece_idx]
angle = (dest_rot_idx - rot_idx) * 90 / 180 * np.pi
# Center coordinates on origin
cx = x - (size - 1) / 2.
cy = y - (size - 1) / 2.
# Perform rotation
nx = np.cos(angle) * cx - np.sin(angle) * cy
ny = np.sin(angle) * cx + np.cos(angle) * cy
# Translate back and round coordinates to _nearest_ integer
nx = nx + (size - 1) / 2.
ny = ny + (size - 1) / 2.
nx = int(np.rint(nx))
ny = int(np.rint(ny))
# Perform swap if piece is an edge, and swap == 1 at NEW location
new_piece_idx = pieces[:,ny,nx].argmax()
edge_idx = new_piece_idx % 4
if new_piece_idx >= 8 and edge_swaps[edge_idx] == 1:
is_f_edge = (new_piece_idx - 8) // 4 # 1 if f, 0 if e edge
edge_type_parity = 1 - 2 * is_f_edge
rotation_parity = 1 - 2 * (edge_idx // 2)
swap_dist = size // 4
# if edge_idx is even, swap in x direction, else y
if edge_idx % 2 == 0:
nx = nx + swap_dist * edge_type_parity * rotation_parity
else:
ny = ny + swap_dist * edge_type_parity * rotation_parity
# append new index to permutation array
new_idx = int(ny * size + nx)
perm.append(new_idx)
# sanity check
#import matplotlib.pyplot as plt
#missing = sorted(set(range(size*size)).difference(set(perm)))
#asdf = np.zeros(size*size)
#asdf[missing] = 1
#plt.imshow(asdf.reshape(size,size))
#plt.savefig('tmp.png')
#plt.show()
#print(np.sum(asdf))
#viz = np.zeros((64,64))
#for idx in perm:
# y, x = idx // 64, idx % 64
# viz[y,x] = 1
#plt.imshow(viz)
#plt.savefig('tmp.png')
#Image.fromarray(viz * 255).convert('RGB').save('tmp.png')
#Image.fromarray(pieces_edge1[0] * 255).convert('RGB').save('tmp.png')
# sanity check on test image
#im = Image.open('results/flip.campfire.man/0000/sample_64.png')
#im = Image.open('results/flip.campfire.man/0000/sample_256.png')
#im = np.array(im)
#Image.fromarray(im.reshape(-1, 3)[perm].reshape(size,size,3)).save('test.png')
return torch.tensor(perm), (piece_perms, edge_swaps)
#for i in range(100):
#make_jigsaw_perm(64, seed=i)
#make_jigsaw_perm(256, seed=11)
def recover_patch_permute(im_0, im_1, patch_size):
'''
Given two views of a patch permutation illusion, recover the patch
permutation used.
im_0 (PIL.Image) :
Identity view of the illusion
im_1 (PIL.Image) :
Patch permuted view of the illusion
patch_size (int) :
Size of the patches in the image
'''
# Convert to tensors
im_0 = TF.to_tensor(im_0)
im_1 = TF.to_tensor(im_1)
# Extract patches
patches_0 = rearrange(im_0,
'c (h p1) (w p2) -> (h w) c p1 p2',
p1=patch_size,
p2=patch_size)
patches_1 = rearrange(im_1,
'c (h p1) (w p2) -> (h w) c p1 p2',
p1=patch_size,
p2=patch_size)
# Repeat patches_1 for each patch in patches_0
patches_1_repeated = repeat(patches_1,
'np c p1 p2 -> np1 np c p1 p2',
np=patches_1.shape[0],
np1=patches_1.shape[0],
p1=patch_size,
p2=patch_size)
# Find closest patch in other image by L1 dist, and return indexes
perm = (patches_1_repeated - patches_0[:,None]).abs().sum((2,3,4)).argmin(1)
return perm
|