def load_db_detect_train_neg():
#neg image cropping
nid = 0
neg_file_list = os.listdir(etc.neg_dir)
print "neg data number={}".format(len(neg_file_list))
neg_db = [0 for n in xrange(len(neg_file_list))]
neg_img = [0 for n in xrange(len(neg_file_list))]
for file in neg_file_list:
img = Image.open(etc.neg_dir + file)
if len(np.shape(np.asarray(img))) != 3:
continue
neg_img[nid] = img
if (nid + 1 % 2000 == 0):
print nid + 1, "/", len(neg_file_list), "th neg image cropping..."
neg_db_line = np.zeros((etc.neg_per_img, etc.dim_12), np.float32)
for neg_iter in xrange(etc.neg_per_img):
rad_rand = randint(0, min(img.size[0], img.size[1]) - 1)
while (rad_rand <= etc.face_minimum):
rad_rand = randint(0, min(img.size[0], img.size[1]) - 1)
x_rand = randint(0, img.size[0] - rad_rand - 1)
y_rand = randint(0, img.size[1] - rad_rand - 1)
neg_cropped_img = img.crop(
(x_rand, y_rand, x_rand + rad_rand, y_rand + rad_rand))
neg_cropped_arr = etc.img2array(neg_cropped_img, etc.img_size_12)
#for debugging
#neg_cropped_img.save(etc.neg_dir + str(f_num) + "_" + str(r) + ".jpg")
neg_db_line[neg_iter, :] = neg_cropped_arr
neg_db[nid] = neg_db_line
nid += 1
neg_db = [elem for elem in neg_db if type(elem) != int]
neg_db = np.vstack(neg_db)
neg_img = [elem for elem in neg_img if type(elem) != int]
print "Neg: " + str(np.shape(neg_db)[0])
return [neg_db, neg_img]
def load_db_detect_train_neg():
#neg image cropping
nid = 0
neg_file_list = os.listdir(etc.neg_dir)
print "neg data number={}".format(len(neg_file_list))
neg_db = [0 for n in xrange(len(neg_file_list))]
neg_img = [0 for n in xrange(len(neg_file_list))]
for file in neg_file_list:
img = Image.open(etc.neg_dir + file)
if len(np.shape(np.asarray(img))) != 3:
continue
neg_img[nid] = img
if(nid+1%2000==0):
print nid+1, "/" , len(neg_file_list), "th neg image cropping..."
neg_db_line = np.zeros((etc.neg_per_img,etc.dim_12), np.float32)
for neg_iter in xrange(etc.neg_per_img):
rad_rand = randint(0,min(img.size[0],img.size[1])-1)
while(rad_rand <= etc.face_minimum):
rad_rand = randint(0,min(img.size[0],img.size[1])-1)
x_rand = randint(0, img.size[0] - rad_rand - 1)
y_rand = randint(0, img.size[1] - rad_rand - 1)
neg_cropped_img = img.crop((x_rand, y_rand, x_rand + rad_rand, y_rand + rad_rand))
neg_cropped_arr = etc.img2array(neg_cropped_img,etc.img_size_12)
#for debugging
#neg_cropped_img.save(etc.neg_dir + str(f_num) + "_" + str(r) + ".jpg")
neg_db_line[neg_iter,:] = neg_cropped_arr
neg_db[nid] = neg_db_line
nid += 1
neg_db = [elem for elem in neg_db if type(elem) != int]
neg_db = np.vstack(neg_db)
neg_img = [elem for elem in neg_img if type(elem) != int]
print "Neg: " + str(np.shape(neg_db)[0])
return [neg_db, neg_img]
for e in xrange(etc.epoch_calib_num[cascade_lv]):
fp_conf_mat = open("./result/conf_mat_cali_" + str(e) + "_.txt","w")
epoch_start = time.time()
average_loss = 0
for b_iter in xrange(etc.batch_iter):
db_id = random.sample(xrange(len(train_db)),etc.mini_batch)
if cascade_lv == 0:
lb[:,:] = 0.0
for id_,did in enumerate(db_id):
patch = etc.img2array(train_db[did][0],etc.img_size_12)
calib_idx = train_db[did][1]
db_12[id_,:] = patch
lb[id_,calib_idx] = 1.0
zipped = zip(db_12, lb)
np.random.shuffle(zipped)
X_12 = [elem[0] for elem in zipped]
Y = [elem[1] for elem in zipped]
X_12 = np.asarray(X_12)
Y = np.reshape(np.asarray(Y),(np.shape(X_12)[0],etc.cali_patt_num))
average_loss += loss_12.eval(feed_dict = {x_12:X_12, y_target:Y})
train_step_12.run(feed_dict = {x_12:X_12, y_target:Y})
elif cascade_lv == 1:
db_12[etc.pos_batch:,:] = neg_db[neg_id,:etc.dim_12]
lb[:etc.pos_batch,:] = np.ones((etc.pos_batch,1),np.float32)
lb[etc.pos_batch:,:] = np.zeros((etc.neg_batch,1),np.float32)
zipped = zip(db_12, lb)
np.random.shuffle(zipped)
X_12 = [elem[0] for elem in zipped]
Y = [elem[1] for elem in zipped]
X_12 = np.asarray(X_12)
Y = np.reshape(np.asarray(Y),(np.shape(X_12)[0],1))
average_loss += loss_12.eval(feed_dict = {x_12:X_12, y_target:Y})
train_step_12.run(feed_dict = {x_12:X_12, y_target:Y})
elif cascade_lv == 1:
db_12[:etc.pos_batch,:] = pos_db[pos_id,:etc.dim_12]
db_24[:etc.pos_batch,:] = pos_db[pos_id,etc.dim_12:etc.dim_12+etc.dim_24]
for eid,id_ in enumerate(neg_id):
neg_12 = etc.img2array(neg_db[id_],etc.img_size_12)
neg_24 = etc.img2array(neg_db[id_],etc.img_size_24)
db_12[etc.pos_batch+eid,:] = neg_12
db_24[etc.pos_batch+eid,:] = neg_24
lb[:etc.pos_batch,:] = np.ones((etc.pos_batch,1),np.float32)
lb[etc.pos_batch:,:] = np.zeros((etc.neg_batch,1),np.float32)
zipped = zip(db_12, db_24,lb)
np.random.shuffle(zipped)
X_12 = [elem[0] for elem in zipped]
X_24 = [elem[1] for elem in zipped]
Y = [elem[2] for elem in zipped]
X_12 = np.asarray(X_12)
X_24 = np.asarray(X_24)
Y = np.reshape(np.asarray(Y),(np.shape(X_12)[0],1))
FROM_12 = h_fc1_12.eval(feed_dict = {x_12:X_12})
if len(detected_list_scale) > 0:
detected_list[scale] = detected_list_scale
detected_list = [elem for elem in detected_list if type(elem) != int]
result_box = [detected_list[i][j] for i in xrange(len(detected_list)) for j in xrange(len(detected_list[i]))]
if len(result_box) > 0:
#12-calib net
test_db_12 = np.zeros((len(result_box),etc.dim_12),np.float32)
for id_,box in enumerate(result_box):
resized_img_12 = etc.img2array(box[5],etc.img_size_12)
test_db_12[id_,:] = resized_img_12
result = h_fc2_12_cali.eval(feed_dict={x_12_cali: test_db_12})
result_box = etc.calib_run(result_box,result,img)
#NMS_fast for each scale
scale_cur = result_box[0][6]
scale_box = []
final_box = []
for id_,box in enumerate(result_box):
if box[6] == scale_cur:
scale_box.append(box)
if box[6] != scale_cur or id_ == len(result_box)-1:
scale_box.sort(key=lambda x :x[4])
X_12 = [elem[0] for elem in zipped]
Y = [elem[1] for elem in zipped]
X_12 = np.asarray(X_12)
Y = np.reshape(np.asarray(Y), (np.shape(X_12)[0], 1))
average_loss += loss_12.eval(feed_dict={
x_12: X_12,
y_target: Y
})
train_step_12.run(feed_dict={x_12: X_12, y_target: Y})
elif cascade_lv == 1:
db_12[:etc.pos_batch, :] = pos_db[pos_id, :etc.dim_12]
db_24[:etc.pos_batch, :] = pos_db[pos_id,
etc.dim_12:etc.dim_12 +
etc.dim_24]
for eid, id_ in enumerate(neg_id):
neg_12 = etc.img2array(neg_db[id_], etc.img_size_12)
neg_24 = etc.img2array(neg_db[id_], etc.img_size_24)
db_12[etc.pos_batch + eid, :] = neg_12
db_24[etc.pos_batch + eid, :] = neg_24
lb[:etc.pos_batch, :] = np.ones((etc.pos_batch, 1),
np.float32)
lb[etc.pos_batch:, :] = np.zeros((etc.neg_batch, 1),
np.float32)
zipped = zip(db_12, db_24, lb)
np.random.shuffle(zipped)
X_12 = [elem[0] for elem in zipped]
X_24 = [elem[1] for elem in zipped]
Y = [elem[2] for elem in zipped]
X_12 = np.asarray(X_12)
X_24 = np.asarray(X_24)
Y = np.reshape(np.asarray(Y), (np.shape(X_12)[0], 1))
arr = np.asarray(img)
img = np.zeros((img.size[1], img.size[0], etc.input_channel), np.uint8)
img[:, :, 0] = arr
img[:, :, 1] = arr
img[:, :, 2] = arr
img = Image.fromarray(img)
#12-net
result_box = etc.slidingW_Test(img, thr_12, x_12, h_fc2_12)
if len(result_box) > 0:
#12-calib net
test_db_12 = np.zeros((len(result_box), etc.dim_12), np.float32)
for id_, box in enumerate(result_box):
resized_img_12 = etc.img2array(box[5], etc.img_size_12)
test_db_12[id_, :] = resized_img_12
result = h_fc2_12_cali.eval(feed_dict={x_12_cali: test_db_12})
result_box = etc.calib_run(result_box, result, img)
#NMS_fast for each scale
scale_cur = result_box[0][6]
scale_box = []
final_box = []
for id_, box in enumerate(result_box):
if box[6] == scale_cur:
scale_box.append(box)
if box[6] != scale_cur or id_ == len(result_box) - 1:
scale_box.sort(key=lambda x: x[4])
scale_box.reverse()
def load_db_detect_train_12_net_pos():
print "Loading training db..."
#annot_dir = etc.db_dir + "aflw/data/"
annot_dir ="/home/gavinpan/workspace/dataset/aflw/data/"
#annot_fp = open(annot_dir + "annot", "r")
annot_fp = open("face_rect.txt", "r")
raw_data = annot_fp.readlines()
annot_fp.close()
#pos image cropping
print "pos data number={}".format(len(raw_data)-1)
pos_db = [0 for _ in xrange(len(raw_data)-1)]
for i,line in enumerate(raw_data):
if(not line.startswith('#')):
if(i%5000==0):
print i, "/", len(raw_data), "th pos image cropping..."
parsed_line = line.split()
imagePath = parsed_line[1].strip()
imagePath = annot_dir + imagePath
img = Image.open(imagePath)
#check if not RGB
if len(np.shape(np.asarray(img))) != 3 or np.shape(np.asarray(img))[2] != etc.input_channel:
continue
#pos_db_line = np.zeros((2,etc.dim_12 + etc.dim_24 + etc.dim_48), np.float32)
pos_db_line = np.zeros((2,etc.dim_12), np.float32)
#left = int(parsed_line[1])
#upper = int(parsed_line[2])
#right = int(parsed_line[3]) + left
#lower = int(parsed_line[4]) + upper
left = int(parsed_line[2])
upper = int(parsed_line[3])
right = int(parsed_line[4]) + left
lower = int(parsed_line[5]) + upper
if right >= img.size[0]:
right = img.size[0]-1
if lower >= img.size[1]:
lower = img.size[1]-1
#crop
cropped_img = img.crop((left, upper, right, lower))
#flip
flipped_img = cropped_img.transpose(Image.FLIP_LEFT_RIGHT)
cropped_arr_12 = etc.img2array(cropped_img,etc.img_size_12)
flipped_arr_12 = etc.img2array(flipped_img,etc.img_size_12)
'''
cropped_arr_24 = etc.img2array(cropped_img,etc.img_size_24)
flipped_arr_24 = etc.img2array(flipped_img,etc.img_size_24)
cropped_arr_48 = etc.img2array(cropped_img,etc.img_size_48)
flipped_arr_48 = etc.img2array(flipped_img,etc.img_size_48)
'''
#for debugging
#cropped_img.save(etc.pos_dir + str(i) + ".jpg")
pos_db_line[0,:etc.dim_12] = cropped_arr_12
'''
pos_db_line[0,etc.dim_12:etc.dim_12+etc.dim_24] = cropped_arr_24
pos_db_line[0,etc.dim_12+etc.dim_24:] = cropped_arr_48
'''
pos_db_line[1,:etc.dim_12] = flipped_arr_12
'''
pos_db_line[1,etc.dim_12:etc.dim_12+etc.dim_24] = flipped_arr_24
pos_db_line[1,etc.dim_12+etc.dim_24:] = flipped_arr_48
'''
pos_db[i-1] = pos_db_line
#end for load raw data
pos_db = [elem for elem in pos_db if type(elem) != int]
pos_db = np.vstack(pos_db)
print "Pos: " + str(np.shape(pos_db)[0])
return pos_db
def load_db_detect_train():
print "Loading training db..."
#annot_dir = etc.db_dir + "aflw/data/"
annot_dir ="/home/gavinpan/workspace/dataset/aflw/data/"
#annot_fp = open(annot_dir + "annot", "r")
annot_fp = open("face_rect.txt", "r")
raw_data = annot_fp.readlines()
annot_fp.close()
#pos image cropping
print "pos data number={}".format(len(raw_data)-1)
pos_db = [0 for _ in xrange(len(raw_data)-1)]
for i,line in enumerate(raw_data):
if(not line.startswith('#')):
if(i%5000==0):
print i, "/", len(raw_data), "th pos image cropping..."
parsed_line = line.split()
imagePath = parsed_line[1].strip()
imagePath = annot_dir + imagePath
img = Image.open(imagePath)
#check if not RGB
if len(np.shape(np.asarray(img))) != 3 or np.shape(np.asarray(img))[2] != etc.input_channel:
continue
pos_db_line = np.zeros((2,etc.dim_12 + etc.dim_24 + etc.dim_48), np.float32)
#left = int(parsed_line[1])
#upper = int(parsed_line[2])
#right = int(parsed_line[3]) + left
#lower = int(parsed_line[4]) + upper
left = int(parsed_line[2])
upper = int(parsed_line[3])
right = int(parsed_line[4]) + left
lower = int(parsed_line[5]) + upper
if right >= img.size[0]:
right = img.size[0]-1
if lower >= img.size[1]:
lower = img.size[1]-1
#crop
cropped_img = img.crop((left, upper, right, lower))
#flip
flipped_img = cropped_img.transpose(Image.FLIP_LEFT_RIGHT)
cropped_arr_12 = etc.img2array(cropped_img,etc.img_size_12)
flipped_arr_12 = etc.img2array(flipped_img,etc.img_size_12)
cropped_arr_24 = etc.img2array(cropped_img,etc.img_size_24)
flipped_arr_24 = etc.img2array(flipped_img,etc.img_size_24)
cropped_arr_48 = etc.img2array(cropped_img,etc.img_size_48)
flipped_arr_48 = etc.img2array(flipped_img,etc.img_size_48)
#for debugging
#cropped_img.save(etc.pos_dir + str(i) + ".jpg")
pos_db_line[0,:etc.dim_12] = cropped_arr_12
pos_db_line[0,etc.dim_12:etc.dim_12+etc.dim_24] = cropped_arr_24
pos_db_line[0,etc.dim_12+etc.dim_24:] = cropped_arr_48
pos_db_line[1,:etc.dim_12] = flipped_arr_12
pos_db_line[1,etc.dim_12:etc.dim_12+etc.dim_24] = flipped_arr_24
pos_db_line[1,etc.dim_12+etc.dim_24:] = flipped_arr_48
pos_db[i-1] = pos_db_line
#end for load raw data
pos_db = [elem for elem in pos_db if type(elem) != int]
pos_db = np.vstack(pos_db)
#neg image cropping
nid = 0
neg_file_list = os.listdir(etc.neg_dir)
neg_db = [0 for n in xrange(len(neg_file_list))]
neg_img = [0 for n in xrange(len(neg_file_list))]
for file in neg_file_list:
img = Image.open(etc.neg_dir + file)
if len(np.shape(np.asarray(img))) != 3:
continue
neg_img[nid] = img
print nid+1, "/" , len(neg_file_list), "th neg image cropping..."
neg_db_line = np.zeros((etc.neg_per_img,etc.dim_12), np.float32)
for neg_iter in xrange(etc.neg_per_img):
rad_rand = randint(0,min(img.size[0],img.size[1])-1)
while(rad_rand <= etc.face_minimum):
rad_rand = randint(0,min(img.size[0],img.size[1])-1)
x_rand = randint(0, img.size[0] - rad_rand - 1)
y_rand = randint(0, img.size[1] - rad_rand - 1)
neg_cropped_img = img.crop((x_rand, y_rand, x_rand + rad_rand, y_rand + rad_rand))
neg_cropped_arr = etc.img2array(neg_cropped_img,etc.img_size_12)
#for debugging
#neg_cropped_img.save(etc.neg_dir + str(f_num) + "_" + str(r) + ".jpg")
neg_db_line[neg_iter,:] = neg_cropped_arr
neg_db[nid] = neg_db_line
nid += 1
neg_db = [elem for elem in neg_db if type(elem) != int]
neg_db = np.vstack(neg_db)
neg_img = [elem for elem in neg_img if type(elem) != int]
print "Pos: " + str(np.shape(pos_db)[0])
print "Neg: " + str(np.shape(neg_db)[0])
return [pos_db, neg_db, neg_img]
def load_db_detect_train():
print "Loading training db..."
annot_dir = etc.db_dir + "AFLW/aflw/data/"
annot_fp = open(annot_dir + "annot", "r")
raw_data = annot_fp.readlines()
#pos image cropping
pos_db = [0 for _ in xrange(len(raw_data))]
for i, line in enumerate(raw_data):
print i, "/", len(raw_data), "th pos image cropping..."
parsed_line = line.split(',')
img = Image.open(etc.pos_dir + parsed_line[0])
#check if not RGB
if len(np.shape(np.asarray(img))) != 3 or np.shape(
np.asarray(img))[2] != etc.input_channel:
continue
pos_db_line = np.zeros((2, etc.dim_12 + etc.dim_24 + etc.dim_48),
np.float32)
left = int(parsed_line[1])
upper = int(parsed_line[2])
right = int(parsed_line[3]) + left
lower = int(parsed_line[4]) + upper
if right >= img.size[0]:
right = img.size[0] - 1
if lower >= img.size[1]:
lower = img.size[1] - 1
cropped_img = img.crop((left, upper, right, lower))
flipped_img = cropped_img.transpose(Image.FLIP_LEFT_RIGHT)
cropped_arr_12 = etc.img2array(cropped_img, etc.img_size_12)
flipped_arr_12 = etc.img2array(flipped_img, etc.img_size_12)
cropped_arr_24 = etc.img2array(cropped_img, etc.img_size_24)
flipped_arr_24 = etc.img2array(flipped_img, etc.img_size_24)
cropped_arr_48 = etc.img2array(cropped_img, etc.img_size_48)
flipped_arr_48 = etc.img2array(flipped_img, etc.img_size_48)
#for debugging
#cropped_img.save(etc.pos_dir + str(i) + ".jpg")
pos_db_line[0, :etc.dim_12] = cropped_arr_12
pos_db_line[0, etc.dim_12:etc.dim_12 + etc.dim_24] = cropped_arr_24
pos_db_line[0, etc.dim_12 + etc.dim_24:] = cropped_arr_48
pos_db_line[1, :etc.dim_12] = flipped_arr_12
pos_db_line[1, etc.dim_12:etc.dim_12 + etc.dim_24] = flipped_arr_24
pos_db_line[1, etc.dim_12 + etc.dim_24:] = flipped_arr_48
pos_db[i] = pos_db_line
pos_db = [elem for elem in pos_db if type(elem) != int]
pos_db = np.vstack(pos_db)
#neg image cropping
nid = 0
neg_file_list = os.listdir(etc.neg_dir)
neg_db = [0 for n in xrange(len(neg_file_list))]
neg_img = [0 for n in xrange(len(neg_file_list))]
for file in neg_file_list:
img = Image.open(etc.neg_dir + file)
if len(np.shape(np.asarray(img))) != 3:
continue
neg_img[nid] = img
print nid + 1, "/", len(neg_file_list), "th neg image cropping..."
neg_db_line = np.zeros((etc.neg_per_img, etc.dim_12), np.float32)
for neg_iter in xrange(etc.neg_per_img):
rad_rand = randint(0, min(img.size[0], img.size[1]) - 1)
while (rad_rand <= etc.face_minimum):
rad_rand = randint(0, min(img.size[0], img.size[1]) - 1)
x_rand = randint(0, img.size[0] - rad_rand - 1)
y_rand = randint(0, img.size[1] - rad_rand - 1)
neg_cropped_img = img.crop(
(x_rand, y_rand, x_rand + rad_rand, y_rand + rad_rand))
neg_cropped_arr = etc.img2array(neg_cropped_img, etc.img_size_12)
#for debugging
#neg_cropped_img.save(etc.neg_dir + str(f_num) + "_" + str(r) + ".jpg")
neg_db_line[neg_iter, :] = neg_cropped_arr
neg_db[nid] = neg_db_line
nid += 1
neg_db = [elem for elem in neg_db if type(elem) != int]
neg_db = np.vstack(neg_db)
neg_img = [elem for elem in neg_img if type(elem) != int]
print "Pos: " + str(np.shape(pos_db)[0])
print "Neg: " + str(np.shape(neg_db)[0])
return [pos_db, neg_db, neg_img]
for e in xrange(etc.epoch_calib_num[cascade_lv]):
fp_conf_mat = open("./result/conf_mat_cali_" + str(e) + "_.txt", "w")
epoch_start = time.time()
average_loss = 0
for b_iter in xrange(etc.batch_iter):
db_id = random.sample(xrange(len(train_db)), etc.mini_batch)
if cascade_lv == 0:
lb[:, :] = 0.0
for id_, did in enumerate(db_id):
patch = etc.img2array(train_db[did][0], etc.img_size_12)
calib_idx = train_db[did][1]
db_12[id_, :] = patch
lb[id_, calib_idx] = 1.0
zipped = zip(db_12, lb)
np.random.shuffle(zipped)
X_12 = [elem[0] for elem in zipped]
Y = [elem[1] for elem in zipped]
X_12 = np.asarray(X_12)
Y = np.reshape(np.asarray(Y),
(np.shape(X_12)[0], etc.cali_patt_num))
average_loss += loss_12.eval(feed_dict={
x_12: X_12,
def load_db_detect_train_12_net_pos():
print "Loading training db..."
#annot_dir = etc.db_dir + "aflw/data/"
annot_dir = "/home/gavinpan/workspace/dataset/aflw/data/"
#annot_fp = open(annot_dir + "annot", "r")
annot_fp = open("face_rect.txt", "r")
raw_data = annot_fp.readlines()
annot_fp.close()
#pos image cropping
print "pos data number={}".format(len(raw_data) - 1)
pos_db = [0 for _ in xrange(len(raw_data) - 1)]
for i, line in enumerate(raw_data):
if (not line.startswith('#')):
if (i % 5000 == 0):
print i, "/", len(raw_data), "th pos image cropping..."
parsed_line = line.split()
imagePath = parsed_line[1].strip()
imagePath = annot_dir + imagePath
img = Image.open(imagePath)
#check if not RGB
if len(np.shape(np.asarray(img))) != 3 or np.shape(
np.asarray(img))[2] != etc.input_channel:
continue
#pos_db_line = np.zeros((2,etc.dim_12 + etc.dim_24 + etc.dim_48), np.float32)
pos_db_line = np.zeros((2, etc.dim_12), np.float32)
#left = int(parsed_line[1])
#upper = int(parsed_line[2])
#right = int(parsed_line[3]) + left
#lower = int(parsed_line[4]) + upper
left = int(parsed_line[2])
upper = int(parsed_line[3])
right = int(parsed_line[4]) + left
lower = int(parsed_line[5]) + upper
if right >= img.size[0]:
right = img.size[0] - 1
if lower >= img.size[1]:
lower = img.size[1] - 1
#crop
cropped_img = img.crop((left, upper, right, lower))
#flip
flipped_img = cropped_img.transpose(Image.FLIP_LEFT_RIGHT)
cropped_arr_12 = etc.img2array(cropped_img, etc.img_size_12)
flipped_arr_12 = etc.img2array(flipped_img, etc.img_size_12)
'''
cropped_arr_24 = etc.img2array(cropped_img,etc.img_size_24)
flipped_arr_24 = etc.img2array(flipped_img,etc.img_size_24)
cropped_arr_48 = etc.img2array(cropped_img,etc.img_size_48)
flipped_arr_48 = etc.img2array(flipped_img,etc.img_size_48)
'''
#for debugging
#cropped_img.save(etc.pos_dir + str(i) + ".jpg")
pos_db_line[0, :etc.dim_12] = cropped_arr_12
'''
pos_db_line[0,etc.dim_12:etc.dim_12+etc.dim_24] = cropped_arr_24
pos_db_line[0,etc.dim_12+etc.dim_24:] = cropped_arr_48
'''
pos_db_line[1, :etc.dim_12] = flipped_arr_12
'''
pos_db_line[1,etc.dim_12:etc.dim_12+etc.dim_24] = flipped_arr_24
pos_db_line[1,etc.dim_12+etc.dim_24:] = flipped_arr_48
'''
pos_db[i - 1] = pos_db_line
#end for load raw data
pos_db = [elem for elem in pos_db if type(elem) != int]
pos_db = np.vstack(pos_db)
print "Pos: " + str(np.shape(pos_db)[0])
return pos_db
def procimg(img):
if len(np.shape(np.asarray(img))) < 3:
arr = np.asarray(img)
img = np.zeros((img.size[1], img.size[0], etc.input_channel), np.uint8)
img[:, :, 0] = arr
img[:, :, 1] = arr
img[:, :, 2] = arr
img = Image.fromarray(img)
#12-net
result_box = etc.slidingW_Test(img, thr_12, x_12, h_fc2_12)
if len(result_box) > 0:
#12-calib net
test_db_12 = np.zeros((len(result_box), etc.dim_12), np.float32)
for id_, box in enumerate(result_box):
resized_img_12 = etc.img2array(box[5], etc.img_size_12)
test_db_12[id_, :] = resized_img_12
result = h_fc2_12_cali.eval(feed_dict={x_12_cali: test_db_12})
result_box = etc.calib_run(result_box, result, img)
#NMS_fast for each scale
scale_cur = result_box[0][6]
scale_box = []
final_box = []
for id_, box in enumerate(result_box):
if box[6] == scale_cur:
scale_box.append(box)
if box[6] != scale_cur or id_ == len(result_box) - 1:
scale_box.sort(key=lambda x: x[4])
scale_box.reverse()
supp_box_id = etc.NMS_fast(scale_box)
final_box += [scale_box[i] for i in supp_box_id]
scale_cur += 1
scale_box = [box]
result_box = final_box
final_box = []
if len(result_box) > 0:
#24-net
test_db_12 = np.zeros((len(result_box), etc.dim_12), np.float32)
test_db_24 = np.zeros((len(result_box), etc.dim_24), np.float32)
for id_, box in enumerate(result_box):
resized_img_12 = etc.img2array(box[5], etc.img_size_12)
resized_img_24 = etc.img2array(box[5], etc.img_size_24)
test_db_12[id_, :] = resized_img_12
test_db_24[id_, :] = resized_img_24
final_box = []
test_db = np.empty((0, etc.dim_24), np.float32)
for tit in xrange(len(test_db_12) / etc.test_bench_num + 1):
if tit == len(test_db_12) / etc.test_bench_num:
test_12 = test_db_12[tit * etc.test_bench_num:, :]
test_24 = test_db_24[tit * etc.test_bench_num:, :]
box_tmp = result_box[tit * etc.test_bench_num:]
else:
test_12 = test_db_12[tit * etc.test_bench_num:(tit + 1) *
etc.test_bench_num, :]
test_24 = test_db_24[tit * etc.test_bench_num:(tit + 1) *
etc.test_bench_num, :]
box_tmp = result_box[tit * etc.test_bench_num:(tit + 1) *
etc.test_bench_num]
result = h_fc2_24.eval(feed_dict={x_12: test_12, x_24: test_24})
result_id = np.where(result > thr_24)[0]
test_db = np.append(test_db, test_24[result_id, :], axis=0)
final_box += [box_tmp[i] for i in result_id]
result_box = final_box
final_box = []
if len(result_box) > 0:
#24-net_calib
test_db_24 = test_db
result = h_fc2_24_cali.eval(feed_dict={x_24_cali: test_db_24})
result_box = etc.calib_run(result_box, result, img)
#NMS_fast for each scale
scale_cur = result_box[0][6]
scale_box = []
final_box = []
for id_, box in enumerate(result_box):
if box[6] == scale_cur:
scale_box.append(box)
if box[6] != scale_cur or id_ == len(result_box) - 1:
scale_box.sort(key=lambda x: x[4])
scale_box.reverse()
supp_box_id = etc.NMS_fast(scale_box)
final_box += [scale_box[i] for i in supp_box_id]
scale_cur += 1
scale_box = [box]
result_box = final_box
final_box = []
test_db_12 = np.zeros((len(result_box), etc.dim_12), np.float32)
test_db_24 = np.zeros((len(result_box), etc.dim_24), np.float32)
test_db_48 = np.zeros((len(result_box), etc.dim_48), np.float32)
for id_, box in enumerate(result_box):
resized_img_12 = etc.img2array(box[5], etc.img_size_12)
resized_img_24 = etc.img2array(box[5], etc.img_size_24)
resized_img_48 = etc.img2array(box[5], etc.img_size_48)
test_db_12[id_, :] = resized_img_12
test_db_24[id_, :] = resized_img_24
test_db_48[id_, :] = resized_img_48
for thr_idx in xrange(etc.thr_num):
#thr_ = 1e-3 + thr_idx * 1e-3
thr_ = thr_idx * 1e-2
final_box = []
if len(result_box) > 0:
#48-net
result_box_copy = []
result_box_copy += [
copy.deepcopy([box[0], box[1], box[2], box[3], box[4]]) +
[box[5].copy()] + copy.deepcopy([box[6], box[7]])
for box in result_box
]
final_box = []
for tit in xrange(len(test_db_12) / etc.test_bench_num + 1):
if tit == len(test_db_12) / etc.test_bench_num:
test_12 = test_db_12[tit * etc.test_bench_num:, :]
test_24 = test_db_24[tit * etc.test_bench_num:, :]
test_48 = test_db_48[tit * etc.test_bench_num:, :]
box_tmp = result_box_copy[tit * etc.test_bench_num:]
else:
test_12 = test_db_12[tit * etc.test_bench_num:(tit + 1) *
etc.test_bench_num, :]
test_24 = test_db_24[tit * etc.test_bench_num:(tit + 1) *
etc.test_bench_num, :]
test_48 = test_db_48[tit * etc.test_bench_num:(tit + 1) *
etc.test_bench_num, :]
box_tmp = result_box_copy[tit *
etc.test_bench_num:(tit + 1) *
etc.test_bench_num]
result = h_fc2_48.eval(feed_dict={
x_12: test_12,
x_24: test_24,
x_48: test_48
})
result_id = np.where(result > thr_)[0]
final_box += [box_tmp[i] for i in result_id]
#global_NMS_fast
final_box.sort(key=lambda x: x[4])
final_box.reverse()
supp_box_id = etc.NMS_fast(final_box)
final_box = [final_box[i] for i in supp_box_id]
if len(final_box) > 0:
#48-net_calib
test_db_48_calib = np.zeros((len(final_box), etc.dim_48),
np.float32)
for id_, box in enumerate(final_box):
resized_img_48 = etc.img2array(box[5], etc.img_size_48)
test_db_48_calib[id_, :] = resized_img_48
result = h_fc2_48_cali.eval(
feed_dict={x_48_cali: test_db_48_calib})
final_box = etc.calib_run(final_box, result, img)
for box in final_box:
h = box[3] - box[1]
box[1] -= 0.2 * h
box[1] = max(0, box[1])
ImageDraw.Draw(img).rectangle((box[0], box[1], box[2], box[3]),
outline=(0, 0, 255))
final_box = []
return img
