def draw_box(img, result):
img_w, img_h = img.size
# calculate the tightest valid boundary for this image
tvb_x1 = 0 # for width
tvb_x2 = 0
tvb_y1 = 0 # for height
tvb_y2 = 0
for i in range(0, len(result)):
rbtx = result[i]["boundingbox"]["tl"]["x"]
rbty = result[i]["boundingbox"]["tl"]["y"]
rbsw = result[i]["boundingbox"]["size"]["width"]
rbsh = result[i]["boundingbox"]["size"]["height"]
tvb_y1 = max(tvb_y1, -(rbty - ABOVE_H - WINDOW_H))
tvb_x1 = max(tvb_x1, -(rbtx + rbsw / 2 - WINDOW_W / 2))
tvb_x2 = max(tvb_x2, rbtx + rbsw / 2 + WINDOW_W / 2 - img_w)
bg = Image.new("RGBA", (int(img_w + tvb_x1 + tvb_x2), int(img_h + tvb_y1)), (255, 255, 255, 255))
bg_w, bg_h = bg.size
offset = (int(tvb_x1), int(tvb_y1))
bg.paste(img, offset)
img = bg
offset = {"x": offset[0], "y": offset[1]}
draw = ImageDraw.Draw(img)
prandom_init()
attrs = []
for i in range(0, len(result)):
attrs.append(classifier.classify(result[i]))
# A algorithm to assure the diversity of attributes:
# if there are multiple same attributes, select one of them,
# change the attributes of the corresponding people to a
# secondary attributes. Repeat.
np = len(result)
ca = []
for i in range(0, np):
ca.append(0)
while True:
"""
print "======================"
for i in range(0, np):
print attrs[i][ca[i]][0]
print "======================="
"""
stop = True
for i in range(0, np):
# get the index of the best attribute
idxi = attrs[i][ca[i]][2]
eq = [i]
for j in range(i + 1, np):
idxj = attrs[j][ca[j]][2]
if idxi == idxj:
eq.append(j)
if len(eq) > 1:
s = int(prandom() * len(eq))
ca[eq[s]] += 1
stop = False
break
if stop:
break
for i in range(0, len(result)):
text_fn = LABEL_FN + str(attrs[i][ca[i]][2]) + ".png"
img = draw_single_box(img, draw, result[i]["boundingbox"], offset, Image.open(text_fn))
return img
def classify(result): age = result['age'] # 0: female, 1: male -> 2: female, 1: male gender = int(2.5 - result['sex']) # a feature-value list for convenience of coding fvl = [] # facevalue is in [0, 1] facevalue = result['beauty'] fvl.append(facevalue) # value for beard is in [0, 1] beard = result['beard'] fvl.append(beard) # value for mustache is in [0, 1] mustache = result['mustache'] fvl.append(mustache) # value for glass is in [0, 1] glass = result['glasses'] fvl.append(glass) # value for sunglass is in [0, 1] sunglass = result['sunglasses'] fvl.append(sunglass) # value for eyeclosed is in [0, 1] eyeclosed = result['eye_closed'] fvl.append(eyeclosed) # value for mouthopen is in [0, 1] mouthopen = result['mouth_open_wide'] fvl.append(mouthopen) # value for each emotion is in [0, 1] emotion = result['emotion'] em_name = ['happy', 'angry', 'sad', 'confused', 'disgust', 'surprised', 'calm'] em_val = [0, 0, 0, 0, 0, 0, 0] for em in emotion: for i in range(0, len(em_name)): if em_name[i] == em: em_val[i] = emotion[em] for i in range(0, len(em_val)): fvl.append(em_val[i]) # value for random is in [0, 5] # a pseudo-random number is generated based on the attribute 'smile' smile = result['smile'] * 100 + 2 for i in range(0, int(smile)): prandom() fvl.append(0) fn = codecs.open(CLASSIFY_FN, "r", "utf-8") mread_init() n = int(mread(fn)) attr = [] for i in range(0, n): attr_name = mread(fn) m = int(mread(fn)) if m == -1: # the pidx in 'classify.txt' starts from 1 pidx = int(mread(fn)) - 1 attr.append((attr_name, attr[pidx][1], i + 1)) continue val = 0 # valid age interval valid_age = 0 for j in range(0, m): l = int(mread(fn)) r = int(mread(fn)) a = float(mread(fn)) b = float(mread(fn)) if not ((l >= 0 and age < l) or (r >= 0 and age > r)): valid_age = 1 val = a * age + b if valid_age == 0: attr.append((attr_name, 0, i + 1)) for j in range(0, 16): mread(fn) continue # gender restriction valid_gender = int(mread(fn)) if valid_gender > 0 and valid_gender != gender: attr.append((attr_name, 0, i + 1)) for j in range(0, 15): mread(fn) continue # computing score for each feature sum = 0 if (i == 2): print val for j in range(0, 15): p = float(mread(fn)) sum += abs(p) val += fvl[j] * p attr.append((attr_name, val, i + 1)) best_attr = attr[0] index = 0 for i in range(1, n): if best_attr[1] < attr[i][1]: best_attr = attr[i] index = i print "----------------------------" attr.sort(key=lambda tup: -tup[1]) ed = n for i in range(1, n): if abs(attr[i][1] - attr[i - 1][1]) >= 0.0001: ed = i - 1 break # select a random attribute from similar attributes rand = int(prandom() * (ed + 1)) temp = attr[0] attr[0] = attr[rand] attr[rand] = temp fn.close() return attr return LABEL_FN + str(index + 1) + ".png"
