def main():
print("test")
im_pa = 'D:/dataset/dataset2014/dataset/dynamicBackground/fountain01/input'
im_ft = 'jpg'
fg_pa = 'D:/projects/lab/fgimgs_fountain01_v34'
fg_ft = 'png'
fs_im, fullfs_im = loadFiles_plus(im_pa, im_ft)
fs_fg, fullfs_fg = loadFiles_plus(fg_pa, fg_ft)
idx = 1148
im = torch.tensor(imageio.imread(fullfs_im[idx]), dtype=torch.float)
fgim = torch.tensor(imageio.imread(fullfs_fg[idx]), dtype=torch.float)
# start here
radius = 3
rate = 0.6
bayfgim = bayesRefine(im, fgim, radius, rate)
bayfgim = bayesRefine(im, bayfgim, radius, rate)
bayfgim = bayesRefine(im, bayfgim, radius, rate)
bayfgim = bayesRefine(im, bayfgim, radius, rate)
print(bayfgim.shape)
# bayesRefine(im, fgim, radius, rate):
plt.figure()
plt.subplot(1, 2, 1)
plt.imshow(fgim.detach().cpu().numpy())
plt.subplot(1, 2, 2)
plt.imshow(bayfgim.detach().cpu().numpy())
plt.show()
def main():
im_pa = 'D:/dataset/dataset2014/dataset/dynamicBackground/fountain01/input'
# im_pa = 'D:/dataset/dataset2014/dataset/dynamicBackground/fountain02/input'
im_ft = 'jpg'
# fg_pa = 'D:/projects/lab/fgimgs_fountain01_v34'
fg_pa = 'D:/projects/lab/fgimgs_fountain01_v49_0'
# fg_pa = 'D:/projects/lab/fountain02_v53'
fg_ft = 'png'
gt_pa = 'D:/dataset/dataset2014/dataset/dynamicBackground/fountain01/groundtruth'
# gt_pa = 'D:/dataset/dataset2014/dataset/dynamicBackground/fountain02/groundtruth'
gt_ft = 'png'
use_cuda = torch.cuda.is_available()
print("------------")
print(use_cuda)
print("------------")
torch.manual_seed(0)
device = torch.device("cuda:0" if use_cuda else "cpu")
fs_im, fullfs_im = loadFiles_plus(im_pa, im_ft)
fs_fg, fullfs_fg = loadFiles_plus(fg_pa, fg_ft)
fs_gt, fullfs_gt = loadFiles_plus(gt_pa, gt_ft)
idx = 1148
idx = 753
idx = 1104
idx = 756
idx = 780
idx = 1099
idx = 1100
idx = 700
idx = 701
idx = 683
idx = 715
# start here
idx = 672
idx = 657
idx = 701
idx = 690
idx = 710
idx = 1156
idx = 1116
idx = 706
idx = 711
idx = 708
idx = 704
radius = 2
rate = 0.8
num = 40
TP_sum = 0
FP_sum = 0
TN_sum = 0
FN_sum = 0
TP_sum_bay = 0
FP_sum_bay = 0
TN_sum_bay = 0
FN_sum_bay = 0
im = torch.tensor( imageio.imread(fullfs_im[idx]), dtype = torch.float )
fgim = torch.tensor( imageio.imread(fullfs_fg[idx]), dtype = torch.float )
gtim = torch.tensor( imageio.imread(fullfs_gt[idx]), dtype = torch.float )
starttime = time.time()
bayfgim = bayesRefine_iterative_gpu(im, fgim, radius, rate, num, device)
# bayfgim = bayesRefine_iterative_gpu(im, bayfgim, 1, 0.6, num, device)
endtime = time.time()
print("total time:", endtime - starttime)
fgim = fgim.detach().cpu().numpy()
gtim = gtim.detach().cpu().numpy()
bayfgim = bayfgim.detach().cpu().numpy()
im = im.detach().cpu().numpy()
TP, FP, TN, FN = evaluation_numpy_entry(np.round(fgim), np.round(gtim))
print("fg entry:", TP, FP, TN, FN)
Re = TP/max((TP + FN), 1)
Pr = TP/max((TP + FP), 1)
Fm = (2*Re*Pr)/max((Pr + Re), 0.0001)
TP_sum += TP
FP_sum += FP
TN_sum += TN
FN_sum += FN
Re_sum = TP_sum/max((TP_sum + FN_sum), 1)
Pr_sum = TP_sum/max((TP_sum + FP_sum), 1)
Fm_sum = (2*Re_sum*Pr_sum)/max((Pr_sum + Re_sum), 0.0001)
# print("Re, Pr, Fm:", Re, Pr, Fm)
TP_bay, FP_bay, TN_bay, FN_bay = evaluation_numpy_entry(np.round(bayfgim), np.round(gtim))
print("bay entry:", TP_bay, FP_bay, TN_bay, FN_bay)
Re_bay = TP_bay/max((TP_bay + FN_bay), 1)
Pr_bay = TP_bay/max((TP_bay + FP_bay), 1)
Fm_bay = (2*Re_bay*Pr_bay)/max((Pr_bay + Re_bay), 0.0001)
TP_sum_bay += TP_bay
FP_sum_bay += FP_bay
TN_sum_bay += TN_bay
FN_sum_bay += FN_bay
Re_sum_bay = TP_sum_bay/max((TP_sum_bay + FN_sum_bay), 1)
Pr_sum_bay = TP_sum_bay/max((TP_sum_bay + FP_sum_bay), 1)
Fm_sum_bay = (2*Re_sum_bay*Pr_sum_bay)/max((Pr_sum_bay + Re_sum_bay), 0.0001)
print("")
print("current: ", Re, Pr, Fm)
print("bay: ", Re_bay, Pr_bay, Fm_bay)
print("accumulated: ", Re_sum, Pr_sum, Fm_sum)
print("accumulated bay:", Re_sum_bay, Pr_sum_bay, Fm_sum_bay)
print("")
# for i in range(640, 1499):
# for i in range(640, 1184):
# idx = i
#
#
# im = torch.tensor( imageio.imread(fullfs_im[idx]), dtype = torch.float )
# fgim = torch.tensor( imageio.imread(fullfs_fg[idx]), dtype = torch.float )
# gtim = torch.tensor( imageio.imread(fullfs_gt[idx]), dtype = torch.float )
#
#
#
# starttime = time.time()
# bayfgim = bayesRefine_iterative_gpu(im, fgim, radius, rate, num, device)
# endtime = time.time()
#
# print("total time:", endtime - starttime)
#
#
#
#
# fgim = fgim.detach().cpu().numpy()
# gtim = gtim.detach().cpu().numpy()
# bayfgim = bayfgim.detach().cpu().numpy()
#
# im = im.detach().cpu().numpy()
#
#
# # print(fgim.shape)
#
# # fgim[:, 0:26] = 0
# TP, FP, TN, FN = evaluation_numpy_entry(np.round(fgim), np.round(gtim))
#
# print("fg entry:", TP, FP, TN, FN)
#
#
# Re = TP/max((TP + FN), 1)
# Pr = TP/max((TP + FP), 1)
# Fm = (2*Re*Pr)/max((Pr + Re), 0.0001)
#
#
# TP_sum += TP
# FP_sum += FP
# TN_sum += TN
# FN_sum += FN
#
# Re_sum = TP_sum/max((TP_sum + FN_sum), 1)
# Pr_sum = TP_sum/max((TP_sum + FP_sum), 1)
# Fm_sum = (2*Re_sum*Pr_sum)/max((Pr_sum + Re_sum), 0.0001)
#
#
#
#
# # print("Re, Pr, Fm:", Re, Pr, Fm)
# # bayfgim[:, 0:26] = 0
# TP_bay, FP_bay, TN_bay, FN_bay = evaluation_numpy_entry(np.round(bayfgim), np.round(gtim))
#
# print("bay entry:", TP_bay, FP_bay, TN_bay, FN_bay)
#
# Re_bay = TP_bay/max((TP_bay + FN_bay), 1)
# Pr_bay = TP_bay/max((TP_bay + FP_bay), 1)
# Fm_bay = (2*Re_bay*Pr_bay)/max((Pr_bay + Re_bay), 0.0001)
#
#
# TP_sum_bay += TP_bay
# FP_sum_bay += FP_bay
# TN_sum_bay += TN_bay
# FN_sum_bay += FN_bay
#
# Re_sum_bay = TP_sum_bay/max((TP_sum_bay + FN_sum_bay), 1)
# Pr_sum_bay = TP_sum_bay/max((TP_sum_bay + FP_sum_bay), 1)
# Fm_sum_bay = (2*Re_sum_bay*Pr_sum_bay)/max((Pr_sum_bay + Re_sum_bay), 0.0001)
#
#
# print("")
# print("i = ", i)
# print("current: ", Re, Pr, Fm)
# print("bay: ", Re_bay, Pr_bay, Fm_bay)
# print("accumulated: ", Re_sum, Pr_sum, Fm_sum)
# print("accumulated bay:", Re_sum_bay, Pr_sum_bay, Fm_sum_bay)
# print("")
#
# print("Re, Pr, Fm:", Re, Pr, Fm)
plt.figure()
plt.subplot(2, 2, 1)
# plt.imshow(fgim.detach().cpu().numpy(), cmap='gray')
plt.imshow(fgim, cmap='gray')
plt.subplot(2, 2, 2)
# plt.imshow(bayfgim.detach().cpu().numpy(), cmap='gray')
plt.imshow(bayfgim, cmap='gray')
plt.subplot(2, 2, 3)
plt.imshow(im/255.0)
plt.subplot(2, 2, 4)
plt.imshow(gtim/255.0)
plt.show()
def main():
im_pa = 'D:/dataset/dataset2014/dataset/dynamicBackground/fountain01/input'
im_ft = 'jpg'
fg_pa = 'D:/projects/lab/fgimgs_fountain01_v34'
fg_ft = 'png'
fs_im, fullfs_im = loadFiles_plus(im_pa, im_ft)
fs_fg, fullfs_fg = loadFiles_plus(fg_pa, fg_ft)
idx = 1148
im = torch.tensor(imageio.imread(fullfs_im[idx]), dtype=torch.float)
fgim = torch.tensor(imageio.imread(fullfs_fg[idx]), dtype=torch.float)
# start here
radius = 3
rate = 0.6
row_im, column_im, byte_im = im.shape
starttime = time.time()
impos = img2pos(im)
lab = rgb2lab(im)
row, column, byte = im.shape
features = torch.cat((im, im[:, :, 0:2]), dim=2)
features[:, :, 0] = lab[:, :, 0]
features[:, :, 1] = lab[:, :, 1]
features[:, :, 2] = lab[:, :, 2]
features[:, :, 3] = impos[:, :, 0]
features[:, :, 4] = impos[:, :, 1]
features_vec = features.reshape(row_im * column_im, 5)
print("features_vec.shape:", features_vec.shape)
min_val, min_idx = torch.min(features_vec, dim=0)
features_vec = features_vec - min_val
min_val, min_idx = torch.min(features_vec, dim=0)
max_val, max_idx = torch.max(features_vec, dim=0)
features_vec = features_vec / max_val
features = features_vec.reshape(row_im, column_im, 5)
pixel_features = features
[row_f, column_f, byte_f] = features.shape
ex_feats = F.pad(features.permute(2, 0, 1).unsqueeze(0),
(radius, radius, radius, radius),
mode='replicate').squeeze()
ex_feats = ex_feats.permute(1, 2, 0)
ex_label = F.pad(fgim.unsqueeze(-1).permute(2, 0, 1).unsqueeze(0),
(radius, radius, radius, radius),
mode='replicate').squeeze()
ex_label = ex_label
ex_image = F.pad(im.permute(2, 0, 1).unsqueeze(0),
(radius, radius, radius, radius),
mode='replicate').squeeze()
ex_image = ex_image.permute(1, 2, 0)
# print(ex_feats.shape)
# print(ex_label.shape)
# print(ex_image.shape)
num_feats = radius * 2 + 1
num_feats = num_feats * num_feats
labs_vec = torch.cat([
ex_label[i:i + row_im, j:j + column_im].reshape(row_im * column_im, 1)
for i in range(radius * 2 + 1) for j in range(radius * 2 + 1)
],
dim=1)
imgs_vec = torch.cat([
ex_feats[i:i + row_im, j:j + column_im].reshape(
row_im * column_im, 5).unsqueeze(-1) for i in range(radius * 2 + 1)
for j in range(radius * 2 + 1)
],
dim=2)
pixel_vec = pixel_features.reshape(row_im * column_im, 5)
# print("total cost:", torch.sum( imgs_vec[:, :, 24] - pixel_vec))
dis_vec = imgs_vec.permute(2, 0, 1) - pixel_vec
# print(dis_vec.shape)
# print(torch.sum(dis_vec[24, :, :]))
P0_num = torch.sum(labs_vec == 0, dim=1).float()
P1_num = torch.sum(labs_vec == 255, dim=1).float()
P0 = P0_num / (P0_num + P1_num)
P1 = P1_num / (P0_num + P1_num)
num1, num2 = labs_vec.shape
idxlabs = labs_vec.expand(5, num1, num2).permute(2, 1, 0)
neg_dis = dis_vec.clone()
pos_dis = dis_vec.clone()
neg_dis[idxlabs == 255] = 0
pos_dis[idxlabs == 0] = 0
neg_dis = neg_dis.sum(dim=0).sum(dim=1)
pos_dis = pos_dis.sum(dim=0).sum(dim=1)
neg_dis = neg_dis / P0_num
pos_dis = pos_dis / P1_num
neg_dis = 1.0 / neg_dis
pos_dis = 1.0 / pos_dis
neg_pra = neg_dis * P0
pos_pra = pos_dis * P1
re_labs = torch.abs(fgim - fgim).reshape(row_im * column_im)
re_labs[pos_pra > neg_pra] = 255
re_labs[P0_num == 0] = 255
re_labs[P1_num == 0] = 0
bayfgim = re_labs.reshape(row_im, column_im)
print(bayfgim.shape)
print(P0_num.shape)
print(P1_num.shape)
print(neg_dis.shape)
print(pos_dis.shape)
plt.figure()
plt.subplot(1, 2, 1)
plt.imshow(fgim.detach().cpu().numpy(), cmap='gray')
plt.subplot(1, 2, 2)
plt.imshow(bayfgim.detach().cpu().numpy(), cmap='gray')
plt.show()
#
# sum_dis = torch.abs(dis_vec).sum(dim = 0).sum(dim = 1)
#
#
#
# idx = 42032
#
# print(neg_dis[:, idx, 0])
# print(pos_dis[:, idx, 0])
#
# print(idxlabs[:, idx, 0])
#
#
#
#
# print(neg_dis.shape)
# print(pos_dis.shape)
#
#
#
# print(dis_vec.shape)
# print(labs_vec.shape)
# print(idxlabs.shape)
# print(dis_vec.shape)
# print(labs_vec.shape)
#
#
# dis_vec = dis_vec.permute(2, 1, 0)
# idx = labs_vec == 0
#
# print("dis_vec.shape:", dis_vec.shape)
#
# dis_vec[:, idx] = 0
#
# print(dis_vec.shape)
# dis_neg = torch.sum(dis_vec[:, idx], dim = 2)
# print(dis_neg.shape)
#
# for i in range(124416):
# print(P0[i])
# print(torch.sum(labs_vec[i,:] == 0))
#
# print("")
# print(labs_vec.shape)
# print(imgs_vec.shape)
# print(pixel_vec.shape)
# for i in range(radius*2 + 1):
# for j in range(radius*2 + 1):
# labs = ex_label[i:i + row_im, j:j + column_im].reshape(row_im*column_im, 1)
# print(labs.shape)
# # print(t.shape)
#
#
# # print(labs)
# print(torch.sum( lab_pad[:, 48] - labs.squeeze()))
endtime = time.time()
print("total time:", endtime - starttime)