def writer_jsons():
for idx, img_name in enumerate(os.listdir(dst_imgs_path)):
print(img_name)
ImgInfo = {}
ImgInfo.update({"img_id": img_name})
img_id = img_name.split('.')[0]
dst_json_name = os.path.join(dst_json_path, img_id + '.json')
if os.path.exists(dst_json_name):
continue
else:
imgPath = os.path.join(dst_imgs_path, img_name)
img = Image.open(imgPath)
size_map = cv.imread(imgPath.replace('images', 'size_map'),
cv.IMREAD_GRAYSCALE)
size_map = torch.from_numpy(size_map)
size_map = F.max_pool2d(size_map[None, None, :, :].float(),
(249, 249), 16, 124)
size_map = F.interpolate(size_map, scale_factor=16).squeeze()
size_map = size_map.numpy()
w, h = img.size
w_size_map, h_size_map = size_map.shape
print('resize', w, h, h_size_map, w_size_map)
if img_id <= '1201':
gt_path = os.path.join(train_path,
'img_' + img_id + '_ann.mat')
ori_imgPath = os.path.join(train_path,
'img_' + img_id + '.jpg')
else:
gt_path = os.path.join(
test_path,
'img_' + str(int(img_id) - 1201).zfill(4) + '_ann.mat')
ori_imgPath = os.path.join(
test_path,
'img_' + str(int(img_id) - 1201).zfill(4) + '.jpg')
gtInf = scio.loadmat(gt_path) # format [ w, h ]
ori_img = Image.open(ori_imgPath)
ori_w, ori_h = ori_img.size
print('ori', ori_w, ori_h)
w_rate, h_rate = w / ori_w, h / ori_h
annPoints = gtInf['annPoints']
annPoints[:, 0] = annPoints[:, 0] * w_rate
annPoints[:, 1] = annPoints[:, 1] * h_rate
annPoints = annPoints.astype(int)
# print(annPoints)
ImgInfo.update({"human_num": len(annPoints)})
center_w, center_h = [], []
xy = []
wide, heiht = [], []
for head in annPoints:
x, y = min(head[0], w - 1), min(head[1], h - 1)
center_w.append(x)
center_h.append(y)
xy.append([int(head[0]), int(head[1])])
if ImgInfo["human_num"] > 4:
dists = euclidean_dist(head[None, :], annPoints)
dists = dists.squeeze()
id = np.argsort(dists)
p1_y, p1_x = min(annPoints[id[1]][1],
h - 1), min(annPoints[id[1]][0], w - 1)
p2_y, p2_x = min(annPoints[id[2]][1],
h - 1), min(annPoints[id[2]][0], w - 1)
p3_y, p3_x = min(annPoints[id[3]][1],
h - 1), min(annPoints[id[3]][0], w - 1)
# print(id)
# import pdb
scale = average_del_min([
size_map[y, x], size_map[p1_y, p1_x],
size_map[p2_y, p2_x], size_map[p3_y, p3_x]
])
scale = max(scale, 4)
else:
scale = max(size_map[y, x], 4)
# print(x,y, scale)
area = np.exp(scale)
length = int(np.sqrt(area))
wide.append(length)
heiht.append(length)
ImgInfo.update({"points": xy})
# new_heads = [[i, j] for [i, j] in zip(new_x, new_y)]
xywh = []
for _, (x, y, x_len,
y_len) in enumerate(zip(center_w, center_h, wide, heiht)):
# print(x,y,x_len,y_len)
x_left_top, y_left_top = max(int(x - x_len / 2),
0), max(int(y - y_len / 2), 0)
x_right_bottom, y_right_bottom = min(
int(x + x_len / 2), w - 1), min(int(y + y_len / 2), h - 1)
xywh.append(
[x_left_top, y_left_top, x_right_bottom, y_right_bottom])
ImgInfo.update({"boxes": xywh})
# print(ImgInfo)
# plot(center_w, center_h, 'g*')
# plt.imshow(img)
# for (x_, y_, w_, h_) in ImgInfo["boxes"]:
# plt.gca().add_patch(plt.Rectangle((x_, y_), w_ - x_, h_ - y_, fill=False, edgecolor='r', linewidth=1))
# plt.show()
with open(dst_json_name, 'w') as f:
json.dump(ImgInfo, f)
f_cen_lst = []
for idx, face in enumerate( faces ):
f_eye_l = face['faceLandmarks']['eyeLeftInner']
f_eye_r = face['faceLandmarks']['eyeRightInner']
f_cen = ( (f_eye_l['x'] + f_eye_r['x'])/2.0, (f_eye_l['y'] + f_eye_r['y'])/2.0 )
f_gen = [ 1 if face['attributes']['gender'] else 0 ][0]
f_age_s = face['attributes']['age']
f_age = [ int(fn.median(a_c)) for a_c in age_cate if f_age_s in a_c ][0]
v_label = v_label_dict[ f_age, f_gen ]
G_q.add_node( idx, label=v_label )
f_cen_lst.append( f_cen )
f_cen_combi = list( itertools.combinations( f_cen_lst, 2 ) )
for a, b in f_cen_combi:
px_dist = fn.euclidean_dist( a, b )
a_idx = f_cen_lst.index( a )
b_idx = f_cen_lst.index( b )
G_q.add_edge( a_idx, b_idx, weight=px_dist )
# Minimum Spanning Tree
G_q_mst = nx.minimum_spanning_tree( G_q )
# Set initial number of Order Distance as 0 For MST edges
for v1, v2 in G_q_mst.edges():
G_q_mst[ v1 ][ v2 ][ 'weight' ] = 0
# Assign Order Distance for All Edges
G_q_od = nx.all_pairs_shortest_path_length( G_q_mst, cutoff=4 )
# Remove the edges visiting itself
# Decrease Order Distances with -1
# Copy MST's Order Distance to Edges Attrs. of Graph G
for v1 in G_q_od.keys():
def generate_masks():
file_list = glob.glob(os.path.join(dst_imgs_path, '*.jpg'))
print(len(file_list))
for idx, img_path in enumerate(file_list):
if idx < -1:
break
img_id = img_path.split('/')[-1].split('.')[0]
img_ori = Image.open(img_path)
w, h = img_ori.size
print(img_id)
print(w, h)
mask_map = np.zeros((h, w), dtype='uint8')
gt_name = os.path.join(dst_json_path, img_id.split('.')[0] + '.json')
with open(gt_name) as f:
ImgInfo = json.load(f)
centroid_list = []
wh_list = []
for id, (w_start, h_start, w_end,
h_end) in enumerate(ImgInfo["boxes"], 0):
centroid_list.append([(w_end + w_start) / 2,
(h_end + h_start) / 2])
wh_list.append(
[max((w_end - w_start) / 2, 3),
max((h_end - h_start) / 2, 3)])
# print(len(centroid_list))
centroids = np.array(centroid_list.copy(), dtype='int')
wh = np.array(wh_list.copy(), dtype='int')
wh[wh > 15] = 15
human_num = ImgInfo["human_num"]
for point in centroids:
point = point[None, :]
dists = euclidean_dist(point, centroids)
dists = dists.squeeze()
id = np.argsort(dists)
for start, first in enumerate(id, 0):
if start > 0 and start < 5:
src_point = point.squeeze()
dst_point = centroids[first]
src_w, src_h = wh[id[0]][0], wh[id[0]][1]
dst_w, dst_h = wh[first][0], wh[first][1]
count = 0
if (src_w + dst_w
) - np.abs(src_point[0] - dst_point[0]) > 0 and (
src_h + dst_h) - np.abs(src_point[1] -
dst_point[1]) > 0:
w_reduce = ((src_w + dst_w) -
np.abs(src_point[0] - dst_point[0])) / 2
h_reduce = ((src_h + dst_h) -
np.abs(src_point[1] - dst_point[1])) / 2
threshold_w, threshold_h = max(
-int(max(src_w - w_reduce, dst_w - w_reduce) / 2.),
-60
), max(
-int(max(src_h - h_reduce, dst_h - h_reduce) / 2.),
-60)
else:
threshold_w, threshold_h = max(
-int(max(src_w, dst_w) / 2.),
-60), max(-int(max(src_h, dst_h) / 2.), -60)
# threshold_w, threshold_h = -5, -5
while (src_w + dst_w) - np.abs(
src_point[0] - dst_point[0]) > threshold_w and (
src_h + dst_h) - np.abs(
src_point[1] - dst_point[1]) > threshold_h:
if (dst_w * dst_h) > (src_w * src_h):
wh[first][0] = max(int(wh[first][0] * 0.9), 2)
wh[first][1] = max(int(wh[first][1] * 0.9), 2)
dst_w, dst_h = wh[first][0], wh[first][1]
else:
wh[id[0]][0] = max(int(wh[id[0]][0] * 0.9), 2)
wh[id[0]][1] = max(int(wh[id[0]][1] * 0.9), 2)
src_w, src_h = wh[id[0]][0], wh[id[0]][1]
if human_num >= 3:
dst_point_ = centroids[id[start + 1]]
dst_w_, dst_h_ = wh[id[start +
1]][0], wh[id[start + 1]][1]
if (dst_w_ * dst_h_) > (src_w * src_h) and (
dst_w_ * dst_h_) > (dst_w * dst_h):
if (src_w + dst_w_) - np.abs(
src_point[0] - dst_point_[0]
) > -3 and (src_h + dst_h_) - np.abs(
src_point[1] - dst_point_[1]) > -3:
wh[id[start + 1]][0] = max(
int(wh[id[start + 1]][0] * 0.9), 2)
wh[id[start + 1]][1] = max(
int(wh[id[start + 1]][1] * 0.9), 2)
count += 1
if count > 40:
break
for (center_w, center_h), (width, height) in zip(centroids, wh):
assert (width > 0 and height > 0)
if (0 < center_w < w) and (0 < center_h < h):
h_start = (center_h - height)
h_end = (center_h + height)
w_start = center_w - width
w_end = center_w + width
#
if h_start < 0:
h_start = 0
if h_end > h:
h_end = h
if w_start < 0:
w_start = 0
if w_end > w:
w_end = w
if cycle:
mask = generate_cycle_mask(height, width)
mask_map[h_start:h_end, w_start:w_end] = mask
else:
mask_map[h_start:h_end, w_start:w_end] = 1
mask_map = mask_map * 255
cv.imwrite(os.path.join(dst_mask_path, img_id + '.png'), mask_map,
[cv.IMWRITE_PNG_BILEVEL, 1])
def generate_masks():
file_list = glob.glob(os.path.join(dst_imgs_path, '*.jpg'))
print(len(file_list))
for idx, img_path in enumerate(file_list):
if idx < -1:
break
img_id = img_path.split('/')[-1].split('.')[0]
img_ori = Image.open(img_path)
w, h = img_ori.size
print(img_id)
print(w, h)
mask_map = np.zeros((h, w), dtype='uint8')
gt_name = os.path.join(dst_json_path, img_id.split('.')[0] + '.json')
with open(gt_name) as f:
ImgInfo = json.load(f)
centroid_list = []
wh_list = []
for id, (w_start, h_start, w_end,
h_end) in enumerate(ImgInfo["boxes"], 0):
centroid_list.append([(w_end + w_start) / 2,
(h_end + h_start) / 2])
wh_list.append(
[max((w_end - w_start) / 2, 3),
max((h_end - h_start) / 2, 3)])
# print(len(centroid_list))
centroids = np.array(centroid_list.copy(), dtype='int')
wh = np.array(wh_list.copy(), dtype='int')
wh[wh > 15] = 15
human_num = ImgInfo["human_num"]
for point in centroids:
point = point[None, :]
dists = euclidean_dist(point, centroids)
dists = dists.squeeze()
id = np.argsort(dists)
for start, first in enumerate(id, 0):
if start > 0 and start < 5:
src_point = point.squeeze()
dst_point = centroids[first]
src_w, src_h = wh[id[0]][0], wh[id[0]][1]
dst_w, dst_h = wh[first][0], wh[first][1]
count = 0
if (src_w + dst_w
) - np.abs(src_point[0] - dst_point[0]) > 0 and (
src_h + dst_h) - np.abs(src_point[1] -
dst_point[1]) > 0:
w_reduce = ((src_w + dst_w) -
np.abs(src_point[0] - dst_point[0])) / 2
h_reduce = ((src_h + dst_h) -
np.abs(src_point[1] - dst_point[1])) / 2
threshold_w, threshold_h = max(
-int(max(src_w - w_reduce, dst_w - w_reduce) / 2.),
-60
), max(
-int(max(src_h - h_reduce, dst_h - h_reduce) / 2.),
-60)
else:
threshold_w, threshold_h = max(
-int(max(src_w, dst_w) / 2.),
-60), max(-int(max(src_h, dst_h) / 2.), -60)
# threshold_w, threshold_h = -5, -5
while (src_w + dst_w) - np.abs(
src_point[0] - dst_point[0]) > threshold_w and (
src_h + dst_h) - np.abs(
src_point[1] - dst_point[1]) > threshold_h:
if (dst_w * dst_h) > (src_w * src_h):
wh[first][0] = max(int(wh[first][0] * 0.9), 2)
wh[first][1] = max(int(wh[first][1] * 0.9), 2)
dst_w, dst_h = wh[first][0], wh[first][1]
else:
wh[id[0]][0] = max(int(wh[id[0]][0] * 0.9), 2)
wh[id[0]][1] = max(int(wh[id[0]][1] * 0.9), 2)
src_w, src_h = wh[id[0]][0], wh[id[0]][1]
if human_num >= 3:
dst_point_ = centroids[id[start + 1]]
dst_w_, dst_h_ = wh[id[start +
1]][0], wh[id[start + 1]][1]
if (dst_w_ * dst_h_) > (src_w * src_h) and (
dst_w_ * dst_h_) > (dst_w * dst_h):
if (src_w + dst_w_) - np.abs(
src_point[0] - dst_point_[0]
) > -3 and (src_h + dst_h_) - np.abs(
src_point[1] - dst_point_[1]) > -3:
wh[id[start + 1]][0] = max(
int(wh[id[start + 1]][0] * 0.9), 2)
wh[id[start + 1]][1] = max(
int(wh[id[start + 1]][1] * 0.9), 2)
count += 1
if count > 40:
break
for (center_w, center_h), (width, height) in zip(centroids, wh):
assert (width > 0 and height > 0)
if (0 < center_w < w) and (0 < center_h < h):
h_start = (center_h - height)
h_end = (center_h + height)
w_start = center_w - width
w_end = center_w + width
#
if h_start < 0:
h_start = 0
if h_end > h:
h_end = h
if w_start < 0:
w_start = 0
if w_end > w:
w_end = w
if cycle:
mask = generate_cycle_mask(height, width)
mask_map[h_start:h_end, w_start:w_end] = mask
else:
mask_map[h_start:h_end, w_start:w_end] = 1
mask_map = mask_map * 255
cv.imwrite(os.path.join(dst_mask_path, img_id + '.png'), mask_map,
[cv.IMWRITE_PNG_BILEVEL, 1])
# plt.imshow(img_ori)
saveImg = plt.gca()
plt.imshow(img_ori)
for a, b in zip(centroid_list, wh_list):
x_, y_, w_, h_ = a[0], a[1], b[0], b[1]
saveImg.add_patch(
plt.Rectangle((x_ - w_, y_ - h_),
2 * w_,
2 * h_,
fill=False,
edgecolor='g',
linewidth=1))
saveImg.axes.get_yaxis().set_visible(False)
saveImg.axes.get_xaxis().set_visible(False)
saveImg.spines['top'].set_visible(False)
saveImg.spines['bottom'].set_visible(False)
saveImg.spines['left'].set_visible(False)
saveImg.spines['right'].set_visible(False)
dst_vis_path = os.path.join(dst_Root, 'box_vis')
if not os.path.exists(dst_vis_path):
os.makedirs(dst_vis_path)
plt.savefig(os.path.join(dst_vis_path, img_id + '.jpg'),
bbox_inches='tight',
pad_inches=0,
dpi=300)
plt.close()
f_cen_lst = []
for idx, face in enumerate(faces):
f_eye_l = face['faceLandmarks']['eyeLeftInner']
f_eye_r = face['faceLandmarks']['eyeRightInner']
f_cen = ((f_eye_l['x'] + f_eye_r['x']) / 2.0,
(f_eye_l['y'] + f_eye_r['y']) / 2.0)
f_gen = [1 if face['attributes']['gender'] else 0][0]
f_age_s = face['attributes']['age']
f_age = [int(fn.median(a_c)) for a_c in age_cate if f_age_s in a_c][0]
v_label = v_label_dict[f_age, f_gen]
G_q.add_node(idx, label=v_label)
f_cen_lst.append(f_cen)
f_cen_combi = list(itertools.combinations(f_cen_lst, 2))
for a, b in f_cen_combi:
px_dist = fn.euclidean_dist(a, b)
a_idx = f_cen_lst.index(a)
b_idx = f_cen_lst.index(b)
G_q.add_edge(a_idx, b_idx, weight=px_dist)
# Minimum Spanning Tree
G_q_mst = nx.minimum_spanning_tree(G_q)
# Set initial number of Order Distance as 0 For MST edges
for v1, v2 in G_q_mst.edges():
G_q_mst[v1][v2]['weight'] = 0
# Assign Order Distance for All Edges
G_q_od = nx.all_pairs_shortest_path_length(G_q_mst, cutoff=4)
# Remove the edges visiting itself
# Decrease Order Distances with -1
# Copy MST's Order Distance to Edges Attrs. of Graph G
for v1 in G_q_od.keys():
def build_facegraph(json_obj):
"""
# Build Face Graph from query image
"""
# create Graph of Query & BoFG
G_q = nx.Graph()
G_bofg = nx.Graph()
# from Query
faces = json_obj
age_range = [(1, 3), (3, 8), (8, 13), (13, 20), (20, 37), (37, 66), (66, 85)]
age_cate = [range(s, e) for s, e in age_range]
v_label_dict = {
(1, 1): 0,
(5, 1): 1,
(10, 1): 2,
(16, 1): 3,
(28, 1): 4,
(51, 1): 5,
(75, 1): 6,
(1, 2): 7,
(5, 2): 8,
(10, 2): 9,
(16, 2): 10,
(28, 2): 11,
(51, 2): 12,
(75, 2): 13,
}
f_cen_lst = []
for idx, face in enumerate(faces):
f_eye_l = face["faceLandmarks"]["eyeLeftInner"]
f_eye_r = face["faceLandmarks"]["eyeRightInner"]
f_cen = ((f_eye_l["x"] + f_eye_r["x"]) / 2.0, (f_eye_l["y"] + f_eye_r["y"]) / 2.0)
f_gen = [1 if face["attributes"]["gender"] else 0][0]
f_age_s = face["attributes"]["age"]
f_age = [int(fn.median(a_c)) for a_c in age_cate if f_age_s in a_c][0]
v_label = v_label_dict[f_age, f_gen]
G_q.add_node(idx, label=v_label)
f_cen_lst.append(f_cen)
f_cen_combi = list(itertools.combinations(f_cen_lst, 2))
for a, b in f_cen_combi:
px_dist = fn.euclidean_dist(a, b)
a_idx = f_cen_lst.index(a)
b_idx = f_cen_lst.index(b)
G_q.add_edge(a_idx, b_idx, weight=px_dist)
# Minimum Spanning Tree
G_q_mst = nx.minimum_spanning_tree(G_q)
# Set initial number of Order Distance as 0 For MST edges
for v1, v2 in G_q_mst.edges():
G_q_mst[v1][v2]["weight"] = 0
# Assign Order Distance for All Edges
G_q_od = nx.all_pairs_shortest_path_length(G_q_mst, cutoff=4)
# Remove the edges visiting itself
# Decrease Order Distances with -1
# Copy MST's Order Distance to Edges Attrs. of Graph G
for v1 in G_q_od.keys():
for v2 in G_q_od[v1].keys():
if v1 == v2: # Remove Nodes self to self
del G_q_od[v1][v2]
else: # Modify Weights of Other Nodes & Assign them to Graph G
# Edge Label
order_dist = G_q_od[v1][v2]
order_dist -= 1
G_q.add_edge(v1, v2, weight=order_dist)
return G_q
def build_facegraph(json_obj):
"""
# Build Face Graph from query image
"""
# create Graph of Query & BoFG
G_q = nx.Graph()
G_bofg = nx.Graph()
# from Query
faces = json_obj
age_range = [(1, 3), (3, 8), (8, 13), (13, 20), (20, 37), (37, 66),
(66, 85)]
age_cate = [range(s, e) for s, e in age_range]
v_label_dict = {
(1, 1): 0,
(5, 1): 1,
(10, 1): 2,
(16, 1): 3,
(28, 1): 4,
(51, 1): 5,
(75, 1): 6,
(1, 2): 7,
(5, 2): 8,
(10, 2): 9,
(16, 2): 10,
(28, 2): 11,
(51, 2): 12,
(75, 2): 13
}
f_cen_lst = []
for idx, face in enumerate(faces):
f_eye_l = face['faceLandmarks']['eyeLeftInner']
f_eye_r = face['faceLandmarks']['eyeRightInner']
f_cen = ((f_eye_l['x'] + f_eye_r['x']) / 2.0,
(f_eye_l['y'] + f_eye_r['y']) / 2.0)
f_gen = [1 if face['attributes']['gender'] else 0][0]
f_age_s = face['attributes']['age']
f_age = [int(fn.median(a_c)) for a_c in age_cate if f_age_s in a_c][0]
v_label = v_label_dict[f_age, f_gen]
G_q.add_node(idx, label=v_label)
f_cen_lst.append(f_cen)
f_cen_combi = list(itertools.combinations(f_cen_lst, 2))
for a, b in f_cen_combi:
px_dist = fn.euclidean_dist(a, b)
a_idx = f_cen_lst.index(a)
b_idx = f_cen_lst.index(b)
G_q.add_edge(a_idx, b_idx, weight=px_dist)
# Minimum Spanning Tree
G_q_mst = nx.minimum_spanning_tree(G_q)
# Set initial number of Order Distance as 0 For MST edges
for v1, v2 in G_q_mst.edges():
G_q_mst[v1][v2]['weight'] = 0
# Assign Order Distance for All Edges
G_q_od = nx.all_pairs_shortest_path_length(G_q_mst, cutoff=4)
# Remove the edges visiting itself
# Decrease Order Distances with -1
# Copy MST's Order Distance to Edges Attrs. of Graph G
for v1 in G_q_od.keys():
for v2 in G_q_od[v1].keys():
if v1 == v2: # Remove Nodes self to self
del G_q_od[v1][v2]
else: # Modify Weights of Other Nodes & Assign them to Graph G
# Edge Label
order_dist = G_q_od[v1][v2]
order_dist -= 1
G_q.add_edge(v1, v2, weight=order_dist)
return G_q