Spaces:
Paused
Paused
Create core.py
Browse files- src/core.py +466 -0
src/core.py
ADDED
|
@@ -0,0 +1,466 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import base64
|
| 2 |
+
import json
|
| 3 |
+
import os
|
| 4 |
+
import re
|
| 5 |
+
import time
|
| 6 |
+
import uuid
|
| 7 |
+
from io import BytesIO
|
| 8 |
+
from pathlib import Path
|
| 9 |
+
import cv2
|
| 10 |
+
|
| 11 |
+
# For inpainting
|
| 12 |
+
|
| 13 |
+
import numpy as np
|
| 14 |
+
import pandas as pd
|
| 15 |
+
import streamlit as st
|
| 16 |
+
from PIL import Image
|
| 17 |
+
from streamlit_drawable_canvas import st_canvas
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
import argparse
|
| 21 |
+
import io
|
| 22 |
+
import multiprocessing
|
| 23 |
+
from typing import Union
|
| 24 |
+
|
| 25 |
+
import torch
|
| 26 |
+
|
| 27 |
+
try:
|
| 28 |
+
torch._C._jit_override_can_fuse_on_cpu(False)
|
| 29 |
+
torch._C._jit_override_can_fuse_on_gpu(False)
|
| 30 |
+
torch._C._jit_set_texpr_fuser_enabled(False)
|
| 31 |
+
torch._C._jit_set_nvfuser_enabled(False)
|
| 32 |
+
except:
|
| 33 |
+
pass
|
| 34 |
+
|
| 35 |
+
from src.helper import (
|
| 36 |
+
download_model,
|
| 37 |
+
load_img,
|
| 38 |
+
norm_img,
|
| 39 |
+
numpy_to_bytes,
|
| 40 |
+
pad_img_to_modulo,
|
| 41 |
+
resize_max_size,
|
| 42 |
+
)
|
| 43 |
+
|
| 44 |
+
NUM_THREADS = str(multiprocessing.cpu_count())
|
| 45 |
+
|
| 46 |
+
os.environ["OMP_NUM_THREADS"] = NUM_THREADS
|
| 47 |
+
os.environ["OPENBLAS_NUM_THREADS"] = NUM_THREADS
|
| 48 |
+
os.environ["MKL_NUM_THREADS"] = NUM_THREADS
|
| 49 |
+
os.environ["VECLIB_MAXIMUM_THREADS"] = NUM_THREADS
|
| 50 |
+
os.environ["NUMEXPR_NUM_THREADS"] = NUM_THREADS
|
| 51 |
+
if os.environ.get("CACHE_DIR"):
|
| 52 |
+
os.environ["TORCH_HOME"] = os.environ["CACHE_DIR"]
|
| 53 |
+
|
| 54 |
+
#BUILD_DIR = os.environ.get("LAMA_CLEANER_BUILD_DIR", "./lama_cleaner/app/build")
|
| 55 |
+
|
| 56 |
+
# For Seam-carving
|
| 57 |
+
|
| 58 |
+
from scipy import ndimage as ndi
|
| 59 |
+
|
| 60 |
+
SEAM_COLOR = np.array([255, 200, 200]) # seam visualization color (BGR)
|
| 61 |
+
SHOULD_DOWNSIZE = True # if True, downsize image for faster carving
|
| 62 |
+
DOWNSIZE_WIDTH = 500 # resized image width if SHOULD_DOWNSIZE is True
|
| 63 |
+
ENERGY_MASK_CONST = 100000.0 # large energy value for protective masking
|
| 64 |
+
MASK_THRESHOLD = 10 # minimum pixel intensity for binary mask
|
| 65 |
+
USE_FORWARD_ENERGY = True # if True, use forward energy algorithm
|
| 66 |
+
|
| 67 |
+
device = torch.device("cpu")
|
| 68 |
+
model_path = "./assets/big-lama.pt"
|
| 69 |
+
model = torch.jit.load(model_path, map_location="cpu")
|
| 70 |
+
model = model.to(device)
|
| 71 |
+
model.eval()
|
| 72 |
+
|
| 73 |
+
|
| 74 |
+
########################################
|
| 75 |
+
# UTILITY CODE
|
| 76 |
+
########################################
|
| 77 |
+
|
| 78 |
+
|
| 79 |
+
def visualize(im, boolmask=None, rotate=False):
|
| 80 |
+
vis = im.astype(np.uint8)
|
| 81 |
+
if boolmask is not None:
|
| 82 |
+
vis[np.where(boolmask == False)] = SEAM_COLOR
|
| 83 |
+
if rotate:
|
| 84 |
+
vis = rotate_image(vis, False)
|
| 85 |
+
cv2.imshow("visualization", vis)
|
| 86 |
+
cv2.waitKey(1)
|
| 87 |
+
return vis
|
| 88 |
+
|
| 89 |
+
def resize(image, width):
|
| 90 |
+
dim = None
|
| 91 |
+
h, w = image.shape[:2]
|
| 92 |
+
dim = (width, int(h * width / float(w)))
|
| 93 |
+
image = image.astype('float32')
|
| 94 |
+
return cv2.resize(image, dim)
|
| 95 |
+
|
| 96 |
+
def rotate_image(image, clockwise):
|
| 97 |
+
k = 1 if clockwise else 3
|
| 98 |
+
return np.rot90(image, k)
|
| 99 |
+
|
| 100 |
+
|
| 101 |
+
########################################
|
| 102 |
+
# ENERGY FUNCTIONS
|
| 103 |
+
########################################
|
| 104 |
+
|
| 105 |
+
def backward_energy(im):
|
| 106 |
+
"""
|
| 107 |
+
Simple gradient magnitude energy map.
|
| 108 |
+
"""
|
| 109 |
+
xgrad = ndi.convolve1d(im, np.array([1, 0, -1]), axis=1, mode='wrap')
|
| 110 |
+
ygrad = ndi.convolve1d(im, np.array([1, 0, -1]), axis=0, mode='wrap')
|
| 111 |
+
|
| 112 |
+
grad_mag = np.sqrt(np.sum(xgrad**2, axis=2) + np.sum(ygrad**2, axis=2))
|
| 113 |
+
|
| 114 |
+
# vis = visualize(grad_mag)
|
| 115 |
+
# cv2.imwrite("backward_energy_demo.jpg", vis)
|
| 116 |
+
|
| 117 |
+
return grad_mag
|
| 118 |
+
|
| 119 |
+
def forward_energy(im):
|
| 120 |
+
"""
|
| 121 |
+
Forward energy algorithm as described in "Improved Seam Carving for Video Retargeting"
|
| 122 |
+
by Rubinstein, Shamir, Avidan.
|
| 123 |
+
Vectorized code adapted from
|
| 124 |
+
https://github.com/axu2/improved-seam-carving.
|
| 125 |
+
"""
|
| 126 |
+
h, w = im.shape[:2]
|
| 127 |
+
im = cv2.cvtColor(im.astype(np.uint8), cv2.COLOR_BGR2GRAY).astype(np.float64)
|
| 128 |
+
|
| 129 |
+
energy = np.zeros((h, w))
|
| 130 |
+
m = np.zeros((h, w))
|
| 131 |
+
|
| 132 |
+
U = np.roll(im, 1, axis=0)
|
| 133 |
+
L = np.roll(im, 1, axis=1)
|
| 134 |
+
R = np.roll(im, -1, axis=1)
|
| 135 |
+
|
| 136 |
+
cU = np.abs(R - L)
|
| 137 |
+
cL = np.abs(U - L) + cU
|
| 138 |
+
cR = np.abs(U - R) + cU
|
| 139 |
+
|
| 140 |
+
for i in range(1, h):
|
| 141 |
+
mU = m[i-1]
|
| 142 |
+
mL = np.roll(mU, 1)
|
| 143 |
+
mR = np.roll(mU, -1)
|
| 144 |
+
|
| 145 |
+
mULR = np.array([mU, mL, mR])
|
| 146 |
+
cULR = np.array([cU[i], cL[i], cR[i]])
|
| 147 |
+
mULR += cULR
|
| 148 |
+
|
| 149 |
+
argmins = np.argmin(mULR, axis=0)
|
| 150 |
+
m[i] = np.choose(argmins, mULR)
|
| 151 |
+
energy[i] = np.choose(argmins, cULR)
|
| 152 |
+
|
| 153 |
+
# vis = visualize(energy)
|
| 154 |
+
# cv2.imwrite("forward_energy_demo.jpg", vis)
|
| 155 |
+
|
| 156 |
+
return energy
|
| 157 |
+
|
| 158 |
+
########################################
|
| 159 |
+
# SEAM HELPER FUNCTIONS
|
| 160 |
+
########################################
|
| 161 |
+
|
| 162 |
+
def add_seam(im, seam_idx):
|
| 163 |
+
"""
|
| 164 |
+
Add a vertical seam to a 3-channel color image at the indices provided
|
| 165 |
+
by averaging the pixels values to the left and right of the seam.
|
| 166 |
+
Code adapted from https://github.com/vivianhylee/seam-carving.
|
| 167 |
+
"""
|
| 168 |
+
h, w = im.shape[:2]
|
| 169 |
+
output = np.zeros((h, w + 1, 3))
|
| 170 |
+
for row in range(h):
|
| 171 |
+
col = seam_idx[row]
|
| 172 |
+
for ch in range(3):
|
| 173 |
+
if col == 0:
|
| 174 |
+
p = np.mean(im[row, col: col + 2, ch])
|
| 175 |
+
output[row, col, ch] = im[row, col, ch]
|
| 176 |
+
output[row, col + 1, ch] = p
|
| 177 |
+
output[row, col + 1:, ch] = im[row, col:, ch]
|
| 178 |
+
else:
|
| 179 |
+
p = np.mean(im[row, col - 1: col + 1, ch])
|
| 180 |
+
output[row, : col, ch] = im[row, : col, ch]
|
| 181 |
+
output[row, col, ch] = p
|
| 182 |
+
output[row, col + 1:, ch] = im[row, col:, ch]
|
| 183 |
+
|
| 184 |
+
return output
|
| 185 |
+
|
| 186 |
+
def add_seam_grayscale(im, seam_idx):
|
| 187 |
+
"""
|
| 188 |
+
Add a vertical seam to a grayscale image at the indices provided
|
| 189 |
+
by averaging the pixels values to the left and right of the seam.
|
| 190 |
+
"""
|
| 191 |
+
h, w = im.shape[:2]
|
| 192 |
+
output = np.zeros((h, w + 1))
|
| 193 |
+
for row in range(h):
|
| 194 |
+
col = seam_idx[row]
|
| 195 |
+
if col == 0:
|
| 196 |
+
p = np.mean(im[row, col: col + 2])
|
| 197 |
+
output[row, col] = im[row, col]
|
| 198 |
+
output[row, col + 1] = p
|
| 199 |
+
output[row, col + 1:] = im[row, col:]
|
| 200 |
+
else:
|
| 201 |
+
p = np.mean(im[row, col - 1: col + 1])
|
| 202 |
+
output[row, : col] = im[row, : col]
|
| 203 |
+
output[row, col] = p
|
| 204 |
+
output[row, col + 1:] = im[row, col:]
|
| 205 |
+
|
| 206 |
+
return output
|
| 207 |
+
|
| 208 |
+
def remove_seam(im, boolmask):
|
| 209 |
+
h, w = im.shape[:2]
|
| 210 |
+
boolmask3c = np.stack([boolmask] * 3, axis=2)
|
| 211 |
+
return im[boolmask3c].reshape((h, w - 1, 3))
|
| 212 |
+
|
| 213 |
+
def remove_seam_grayscale(im, boolmask):
|
| 214 |
+
h, w = im.shape[:2]
|
| 215 |
+
return im[boolmask].reshape((h, w - 1))
|
| 216 |
+
|
| 217 |
+
def get_minimum_seam(im, mask=None, remove_mask=None):
|
| 218 |
+
"""
|
| 219 |
+
DP algorithm for finding the seam of minimum energy. Code adapted from
|
| 220 |
+
https://karthikkaranth.me/blog/implementing-seam-carving-with-python/
|
| 221 |
+
"""
|
| 222 |
+
h, w = im.shape[:2]
|
| 223 |
+
energyfn = forward_energy if USE_FORWARD_ENERGY else backward_energy
|
| 224 |
+
M = energyfn(im)
|
| 225 |
+
|
| 226 |
+
if mask is not None:
|
| 227 |
+
M[np.where(mask > MASK_THRESHOLD)] = ENERGY_MASK_CONST
|
| 228 |
+
|
| 229 |
+
# give removal mask priority over protective mask by using larger negative value
|
| 230 |
+
if remove_mask is not None:
|
| 231 |
+
M[np.where(remove_mask > MASK_THRESHOLD)] = -ENERGY_MASK_CONST * 100
|
| 232 |
+
|
| 233 |
+
seam_idx, boolmask = compute_shortest_path(M, im, h, w)
|
| 234 |
+
|
| 235 |
+
return np.array(seam_idx), boolmask
|
| 236 |
+
|
| 237 |
+
def compute_shortest_path(M, im, h, w):
|
| 238 |
+
backtrack = np.zeros_like(M, dtype=np.int_)
|
| 239 |
+
|
| 240 |
+
|
| 241 |
+
# populate DP matrix
|
| 242 |
+
for i in range(1, h):
|
| 243 |
+
for j in range(0, w):
|
| 244 |
+
if j == 0:
|
| 245 |
+
idx = np.argmin(M[i - 1, j:j + 2])
|
| 246 |
+
backtrack[i, j] = idx + j
|
| 247 |
+
min_energy = M[i-1, idx + j]
|
| 248 |
+
else:
|
| 249 |
+
idx = np.argmin(M[i - 1, j - 1:j + 2])
|
| 250 |
+
backtrack[i, j] = idx + j - 1
|
| 251 |
+
min_energy = M[i - 1, idx + j - 1]
|
| 252 |
+
|
| 253 |
+
M[i, j] += min_energy
|
| 254 |
+
|
| 255 |
+
# backtrack to find path
|
| 256 |
+
seam_idx = []
|
| 257 |
+
boolmask = np.ones((h, w), dtype=np.bool_)
|
| 258 |
+
j = np.argmin(M[-1])
|
| 259 |
+
for i in range(h-1, -1, -1):
|
| 260 |
+
boolmask[i, j] = False
|
| 261 |
+
seam_idx.append(j)
|
| 262 |
+
j = backtrack[i, j]
|
| 263 |
+
|
| 264 |
+
seam_idx.reverse()
|
| 265 |
+
return seam_idx, boolmask
|
| 266 |
+
|
| 267 |
+
########################################
|
| 268 |
+
# MAIN ALGORITHM
|
| 269 |
+
########################################
|
| 270 |
+
|
| 271 |
+
def seams_removal(im, num_remove, mask=None, vis=False, rot=False):
|
| 272 |
+
for _ in range(num_remove):
|
| 273 |
+
seam_idx, boolmask = get_minimum_seam(im, mask)
|
| 274 |
+
if vis:
|
| 275 |
+
visualize(im, boolmask, rotate=rot)
|
| 276 |
+
im = remove_seam(im, boolmask)
|
| 277 |
+
if mask is not None:
|
| 278 |
+
mask = remove_seam_grayscale(mask, boolmask)
|
| 279 |
+
return im, mask
|
| 280 |
+
|
| 281 |
+
|
| 282 |
+
def seams_insertion(im, num_add, mask=None, vis=False, rot=False):
|
| 283 |
+
seams_record = []
|
| 284 |
+
temp_im = im.copy()
|
| 285 |
+
temp_mask = mask.copy() if mask is not None else None
|
| 286 |
+
|
| 287 |
+
for _ in range(num_add):
|
| 288 |
+
seam_idx, boolmask = get_minimum_seam(temp_im, temp_mask)
|
| 289 |
+
if vis:
|
| 290 |
+
visualize(temp_im, boolmask, rotate=rot)
|
| 291 |
+
|
| 292 |
+
seams_record.append(seam_idx)
|
| 293 |
+
temp_im = remove_seam(temp_im, boolmask)
|
| 294 |
+
if temp_mask is not None:
|
| 295 |
+
temp_mask = remove_seam_grayscale(temp_mask, boolmask)
|
| 296 |
+
|
| 297 |
+
seams_record.reverse()
|
| 298 |
+
|
| 299 |
+
for _ in range(num_add):
|
| 300 |
+
seam = seams_record.pop()
|
| 301 |
+
im = add_seam(im, seam)
|
| 302 |
+
if vis:
|
| 303 |
+
visualize(im, rotate=rot)
|
| 304 |
+
if mask is not None:
|
| 305 |
+
mask = add_seam_grayscale(mask, seam)
|
| 306 |
+
|
| 307 |
+
# update the remaining seam indices
|
| 308 |
+
for remaining_seam in seams_record:
|
| 309 |
+
remaining_seam[np.where(remaining_seam >= seam)] += 2
|
| 310 |
+
|
| 311 |
+
return im, mask
|
| 312 |
+
|
| 313 |
+
########################################
|
| 314 |
+
# MAIN DRIVER FUNCTIONS
|
| 315 |
+
########################################
|
| 316 |
+
|
| 317 |
+
def seam_carve(im, dy, dx, mask=None, vis=False):
|
| 318 |
+
im = im.astype(np.float64)
|
| 319 |
+
h, w = im.shape[:2]
|
| 320 |
+
assert h + dy > 0 and w + dx > 0 and dy <= h and dx <= w
|
| 321 |
+
|
| 322 |
+
if mask is not None:
|
| 323 |
+
mask = mask.astype(np.float64)
|
| 324 |
+
|
| 325 |
+
output = im
|
| 326 |
+
|
| 327 |
+
if dx < 0:
|
| 328 |
+
output, mask = seams_removal(output, -dx, mask, vis)
|
| 329 |
+
|
| 330 |
+
elif dx > 0:
|
| 331 |
+
output, mask = seams_insertion(output, dx, mask, vis)
|
| 332 |
+
|
| 333 |
+
if dy < 0:
|
| 334 |
+
output = rotate_image(output, True)
|
| 335 |
+
if mask is not None:
|
| 336 |
+
mask = rotate_image(mask, True)
|
| 337 |
+
output, mask = seams_removal(output, -dy, mask, vis, rot=True)
|
| 338 |
+
output = rotate_image(output, False)
|
| 339 |
+
|
| 340 |
+
elif dy > 0:
|
| 341 |
+
output = rotate_image(output, True)
|
| 342 |
+
if mask is not None:
|
| 343 |
+
mask = rotate_image(mask, True)
|
| 344 |
+
output, mask = seams_insertion(output, dy, mask, vis, rot=True)
|
| 345 |
+
output = rotate_image(output, False)
|
| 346 |
+
|
| 347 |
+
return output
|
| 348 |
+
|
| 349 |
+
|
| 350 |
+
def object_removal(im, rmask, mask=None, vis=False, horizontal_removal=False):
|
| 351 |
+
im = im.astype(np.float64)
|
| 352 |
+
rmask = rmask.astype(np.float64)
|
| 353 |
+
if mask is not None:
|
| 354 |
+
mask = mask.astype(np.float64)
|
| 355 |
+
output = im
|
| 356 |
+
|
| 357 |
+
h, w = im.shape[:2]
|
| 358 |
+
|
| 359 |
+
if horizontal_removal:
|
| 360 |
+
output = rotate_image(output, True)
|
| 361 |
+
rmask = rotate_image(rmask, True)
|
| 362 |
+
if mask is not None:
|
| 363 |
+
mask = rotate_image(mask, True)
|
| 364 |
+
|
| 365 |
+
while len(np.where(rmask > MASK_THRESHOLD)[0]) > 0:
|
| 366 |
+
seam_idx, boolmask = get_minimum_seam(output, mask, rmask)
|
| 367 |
+
if vis:
|
| 368 |
+
visualize(output, boolmask, rotate=horizontal_removal)
|
| 369 |
+
output = remove_seam(output, boolmask)
|
| 370 |
+
rmask = remove_seam_grayscale(rmask, boolmask)
|
| 371 |
+
if mask is not None:
|
| 372 |
+
mask = remove_seam_grayscale(mask, boolmask)
|
| 373 |
+
|
| 374 |
+
num_add = (h if horizontal_removal else w) - output.shape[1]
|
| 375 |
+
output, mask = seams_insertion(output, num_add, mask, vis, rot=horizontal_removal)
|
| 376 |
+
if horizontal_removal:
|
| 377 |
+
output = rotate_image(output, False)
|
| 378 |
+
|
| 379 |
+
return output
|
| 380 |
+
|
| 381 |
+
|
| 382 |
+
|
| 383 |
+
def s_image(im,mask,vs,hs,mode="resize"):
|
| 384 |
+
im = cv2.cvtColor(im, cv2.COLOR_RGBA2RGB)
|
| 385 |
+
mask = 255-mask[:,:,3]
|
| 386 |
+
h, w = im.shape[:2]
|
| 387 |
+
if SHOULD_DOWNSIZE and w > DOWNSIZE_WIDTH:
|
| 388 |
+
im = resize(im, width=DOWNSIZE_WIDTH)
|
| 389 |
+
if mask is not None:
|
| 390 |
+
mask = resize(mask, width=DOWNSIZE_WIDTH)
|
| 391 |
+
|
| 392 |
+
# image resize mode
|
| 393 |
+
if mode=="resize":
|
| 394 |
+
dy = hs#reverse
|
| 395 |
+
dx = vs#reverse
|
| 396 |
+
assert dy is not None and dx is not None
|
| 397 |
+
output = seam_carve(im, dy, dx, mask, False)
|
| 398 |
+
|
| 399 |
+
|
| 400 |
+
# object removal mode
|
| 401 |
+
elif mode=="remove":
|
| 402 |
+
assert mask is not None
|
| 403 |
+
output = object_removal(im, mask, None, False, True)
|
| 404 |
+
|
| 405 |
+
return output
|
| 406 |
+
|
| 407 |
+
|
| 408 |
+
##### Inpainting helper code
|
| 409 |
+
|
| 410 |
+
def run(image, mask):
|
| 411 |
+
"""
|
| 412 |
+
image: [C, H, W]
|
| 413 |
+
mask: [1, H, W]
|
| 414 |
+
return: BGR IMAGE
|
| 415 |
+
"""
|
| 416 |
+
origin_height, origin_width = image.shape[1:]
|
| 417 |
+
image = pad_img_to_modulo(image, mod=8)
|
| 418 |
+
mask = pad_img_to_modulo(mask, mod=8)
|
| 419 |
+
|
| 420 |
+
mask = (mask > 0) * 1
|
| 421 |
+
image = torch.from_numpy(image).unsqueeze(0).to(device)
|
| 422 |
+
mask = torch.from_numpy(mask).unsqueeze(0).to(device)
|
| 423 |
+
|
| 424 |
+
start = time.time()
|
| 425 |
+
with torch.no_grad():
|
| 426 |
+
inpainted_image = model(image, mask)
|
| 427 |
+
|
| 428 |
+
print(f"process time: {(time.time() - start)*1000}ms")
|
| 429 |
+
cur_res = inpainted_image[0].permute(1, 2, 0).detach().cpu().numpy()
|
| 430 |
+
cur_res = cur_res[0:origin_height, 0:origin_width, :]
|
| 431 |
+
cur_res = np.clip(cur_res * 255, 0, 255).astype("uint8")
|
| 432 |
+
cur_res = cv2.cvtColor(cur_res, cv2.COLOR_BGR2RGB)
|
| 433 |
+
return cur_res
|
| 434 |
+
|
| 435 |
+
|
| 436 |
+
def get_args_parser():
|
| 437 |
+
parser = argparse.ArgumentParser()
|
| 438 |
+
parser.add_argument("--port", default=8080, type=int)
|
| 439 |
+
parser.add_argument("--device", default="cuda", type=str)
|
| 440 |
+
parser.add_argument("--debug", action="store_true")
|
| 441 |
+
return parser.parse_args()
|
| 442 |
+
|
| 443 |
+
|
| 444 |
+
def process_inpaint(image, mask):
|
| 445 |
+
image = cv2.cvtColor(image, cv2.COLOR_RGBA2RGB)
|
| 446 |
+
original_shape = image.shape
|
| 447 |
+
interpolation = cv2.INTER_CUBIC
|
| 448 |
+
|
| 449 |
+
#size_limit: Union[int, str] = request.form.get("sizeLimit", "1080")
|
| 450 |
+
#if size_limit == "Original":
|
| 451 |
+
size_limit = max(image.shape)
|
| 452 |
+
#else:
|
| 453 |
+
# size_limit = int(size_limit)
|
| 454 |
+
|
| 455 |
+
print(f"Origin image shape: {original_shape}")
|
| 456 |
+
image = resize_max_size(image, size_limit=size_limit, interpolation=interpolation)
|
| 457 |
+
print(f"Resized image shape: {image.shape}")
|
| 458 |
+
image = norm_img(image)
|
| 459 |
+
|
| 460 |
+
mask = 255-mask[:,:,3]
|
| 461 |
+
mask = resize_max_size(mask, size_limit=size_limit, interpolation=interpolation)
|
| 462 |
+
mask = norm_img(mask)
|
| 463 |
+
|
| 464 |
+
res_np_img = run(image, mask)
|
| 465 |
+
|
| 466 |
+
return cv2.cvtColor(res_np_img, cv2.COLOR_BGR2RGB)
|