Spaces:
Running
Running
File size: 10,466 Bytes
f0de4e8 |
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 |
import numpy as np
import cv2
import os
import tqdm
import time
import torch
import torch.nn as nn
import torch.nn.functional as F
from .util import rgb_to_lab, lab_to_rgb
def blend(f, b, a):
return f*a + b*(1 - a)
class PatchedHarmonizer(nn.Module):
def __init__(self, grid_count=1, init_weights=[0.9, 0.1]):
super(PatchedHarmonizer, self).__init__()
self.eps = 1e-8
# self.weights = torch.nn.Parameter(torch.ones((grid_count, grid_count)), requires_grad=True)
# self.grid_weights_ = torch.nn.Parameter(torch.FloatTensor(init_weights), requires_grad=True)
self.grid_weights = torch.nn.Parameter(
torch.FloatTensor(init_weights), requires_grad=True)
# self.weights.retain_graph = True
self.grid_count = grid_count
def lab_shift(self, x, invert=False):
x = x.float()
if invert:
x[:, 0, :, :] /= 2.55
x[:, 1, :, :] -= 128
x[:, 2, :, :] -= 128
else:
x[:, 0, :, :] *= 2.55
x[:, 1, :, :] += 128
x[:, 2, :, :] += 128
return x
def get_mean_std(self, img, mask, dim=[2, 3]):
sum = torch.sum(img*mask, dim=dim) # (B, C)
num = torch.sum(mask, dim=dim) # (B, C)
mu = sum / (num + self.eps)
mean = mu[:, :, None, None]
var = torch.sum(((img - mean)*mask) ** 2, dim=dim) / (num + self.eps)
var = var[:, :, None, None]
return mean, torch.sqrt(var+self.eps)
def compute_patch_statistics(self, lab):
means, stds = [], []
bs, dx, dy = lab.shape[0], lab.shape[2] // self.grid_count, lab.shape[3] // self.grid_count
for h in range(self.grid_count):
cmeans, cstds = [], []
for w in range(self.grid_count):
ind = [h*dx, (h+1)*dx, w*dy, (w+1)*dy]
if h == self.grid_count - 1:
ind[1] = None
if w == self.grid_count - 1:
ind[-1] = None
m, v = self.compute_mean_var(
lab[:, :, ind[0]:ind[1], ind[2]:ind[3]], dim=[2, 3])
cmeans.append(m)
cstds.append(v)
means.append(cmeans)
stds.append(cstds)
return means, stds
def compute_mean_var(self, x, dim=[1, 2]):
mean = x.float().mean(dim=dim)[:, :, None, None]
var = torch.sqrt(x.float().var(dim=dim))[:, :, None, None]
return mean, var
def forward(self, fg_rgb, bg_rgb, alpha, masked_stats=False):
bg_rgb = F.interpolate(bg_rgb, size=(
fg_rgb.shape[2:])) # b x C x H x W
bg_lab = bg_rgb # self.lab_shift(rgb_to_lab(bg_rgb/255.))
fg_lab = fg_rgb # self.lab_shift(rgb_to_lab(fg_rgb/255.))
if masked_stats:
self.bg_global_mean, self.bg_global_var = self.get_mean_std(
img=bg_lab, mask=(1-alpha))
self.fg_global_mean, self.fg_global_var = self.get_mean_std(
img=fg_lab, mask=torch.ones_like(alpha))
else:
self.bg_global_mean, self.bg_global_var = self.compute_mean_var(bg_lab, dim=[
2, 3])
self.fg_global_mean, self.fg_global_var = self.compute_mean_var(fg_lab, dim=[
2, 3])
self.bg_means, self.bg_vars = self.compute_patch_statistics(
bg_lab)
self.fg_means, self.fg_vars = self.compute_patch_statistics(
fg_lab)
fg_harm = self.harmonize(fg_lab)
# fg_harm = lab_to_rgb(fg_harm)
bg = F.interpolate(bg_rgb, size=(fg_rgb.shape[2:]))/255.
composite = blend(fg_harm, bg, alpha)
return composite, fg_harm
def harmonize(self, fg):
harmonized = torch.zeros_like(fg)
dx = fg.shape[2] // self.grid_count
dy = fg.shape[3] // self.grid_count
for h in range(self.grid_count):
for w in range(self.grid_count):
ind = [h*dx, (h+1)*dx, w*dy, (w+1)*dy]
if h == self.grid_count - 1:
ind[1] = None
if w == self.grid_count - 1:
ind[-1] = None
harmonized[:, :, ind[0]:ind[1], ind[2]:ind[3]] = self.normalize_channel(
fg[:, :, ind[0]:ind[1], ind[2]:ind[3]], h, w)
# harmonized = self.lab_shift(harmonized, invert=True)
return harmonized
def normalize_channel(self, value, h, w):
fg_local_mean, fg_local_var = self.fg_means[h][w], self.fg_vars[h][w]
bg_local_mean, bg_local_var = self.bg_means[h][w], self.bg_vars[h][w]
fg_global_mean, fg_global_var = self.fg_global_mean, self.fg_global_var
bg_global_mean, bg_global_var = self.bg_global_mean, self.bg_global_var
# global2global normalization
zeroed_mean = value - fg_global_mean
# (fg_v * div_global_v + (1-fg_v) * div_v)
scaled_var = zeroed_mean * (bg_global_var/(fg_global_var + self.eps))
normalized_global = scaled_var + bg_global_mean
# local2local normalization
zeroed_mean = value - fg_local_mean
# (fg_v * div_global_v + (1-fg_v) * div_v)
scaled_var = zeroed_mean * (bg_local_var/(fg_local_var + self.eps))
normalized_local = scaled_var + bg_local_mean
return self.grid_weights[0]*normalized_local + self.grid_weights[1]*normalized_global
def normalize_fg(self, value):
zeroed_mean = value - \
(self.fg_local_mean *
self.grid_weights[None, None, :, :, None, None]).sum().squeeze()
# (fg_v * div_global_v + (1-fg_v) * div_v)
scaled_var = zeroed_mean * \
(self.bg_global_var/(self.fg_global_var + self.eps))
normalized_lg = scaled_var + \
(self.bg_local_mean *
self.grid_weights[None, None, :, :, None, None]).sum().squeeze()
return normalized_lg
class PatchNormalizer(nn.Module):
def __init__(self, in_channels=3, eps=1e-7, grid_count=1, weights=[0.5, 0.5], init_value=1e-2):
super(PatchNormalizer, self).__init__()
self.grid_count = grid_count
self.eps = eps
self.weights = nn.Parameter(
torch.FloatTensor(weights), requires_grad=True)
self.fg_var = nn.Parameter(
init_value * torch.ones(in_channels)[None, :, None, None], requires_grad=True)
self.fg_bias = nn.Parameter(
init_value * torch.zeros(in_channels)[None, :, None, None], requires_grad=True)
self.patched_fg_var = nn.Parameter(
init_value * torch.ones(in_channels)[None, :, None, None], requires_grad=True)
self.patched_fg_bias = nn.Parameter(
init_value * torch.zeros(in_channels)[None, :, None, None], requires_grad=True)
self.bg_var = nn.Parameter(
init_value * torch.ones(in_channels)[None, :, None, None], requires_grad=True)
self.bg_bias = nn.Parameter(
init_value * torch.zeros(in_channels)[None, :, None, None], requires_grad=True)
self.grid_weights = torch.nn.Parameter(torch.ones((in_channels, grid_count, grid_count))[
None, :, :, :] / (grid_count*grid_count*in_channels), requires_grad=True)
def local_normalization(self, value):
zeroed_mean = value - \
(self.fg_local_mean *
self.grid_weights[None, None, :, :, None, None]).sum().squeeze()
# (fg_v * div_global_v + (1-fg_v) * div_v)
scaled_var = zeroed_mean * \
(self.bg_global_var/(self.fg_global_var + self.eps))
normalized_lg = scaled_var + \
(self.bg_local_mean *
self.grid_weights[None, None, :, :, None, None]).sum().squeeze()
return normalized_lg
def get_mean_std(self, img, mask, dim=[2, 3]):
sum = torch.sum(img*mask, dim=dim) # (B, C)
num = torch.sum(mask, dim=dim) # (B, C)
mu = sum / (num + self.eps)
mean = mu[:, :, None, None]
var = torch.sum(((img - mean)*mask) ** 2, dim=dim) / (num + self.eps)
var = var[:, :, None, None]
return mean, torch.sqrt(var+self.eps)
def compute_patch_statistics(self, img, mask):
means, stds = [], []
bs, dx, dy = img.shape[0], img.shape[2] // self.grid_count, img.shape[3] // self.grid_count
for h in range(self.grid_count):
cmeans, cstds = [], []
for w in range(self.grid_count):
ind = [h*dx, (h+1)*dx, w*dy, (w+1)*dy]
if h == self.grid_count - 1:
ind[1] = None
if w == self.grid_count - 1:
ind[-1] = None
m, v = self.get_mean_std(
img[:, :, ind[0]:ind[1], ind[2]:ind[3]], mask[:, :, ind[0]:ind[1], ind[2]:ind[3]], dim=[2, 3])
cmeans.append(m.reshape(m.shape[:2]))
cstds.append(v.reshape(v.shape[:2]))
means.append(torch.stack(cmeans))
stds.append(torch.stack(cstds))
return torch.stack(means), torch.stack(stds)
def compute_mean_var(self, x, dim=[2, 3]):
mean = x.float().mean(dim=dim)
var = torch.sqrt(x.float().var(dim=dim))
return mean, var
def forward(self, fg, bg, mask):
self.local_means, self.local_vars = self.compute_patch_statistics(
bg, (1-mask))
bg_mean, bg_var = self.get_mean_std(bg, 1 - mask)
zeroed_mean = (bg - bg_mean)
unscaled = zeroed_mean / bg_var
bg_normalized = unscaled * self.bg_var + self.bg_bias
fg_mean, fg_var = self.get_mean_std(fg, mask)
zeroed_mean = fg - fg_mean
unscaled = zeroed_mean / fg_var
mean_patched_back = (self.local_means.permute(
2, 3, 0, 1)*self.grid_weights).sum(dim=[2, 3])[:, :, None, None]
normalized = unscaled * bg_var + bg_mean
patch_normalized = unscaled * bg_var + mean_patched_back
fg_normalized = normalized * self.fg_var + self.fg_bias
fg_patch_normalized = patch_normalized * \
self.patched_fg_var + self.patched_fg_bias
fg_result = self.weights[0] * fg_normalized + \
self.weights[1] * fg_patch_normalized
composite = blend(fg_result, bg_normalized, mask)
return composite
|