def gen_positive_list_voc_format():
'''
out_file = "/ssd/xulifeng/workspace/hd_densebox/train_data/fid_fullframe_lst.txt"
dataset_list = ['/ssd/xulifeng/train_data/head_detection/fid/HeadVocFormat/FID_DID_HEAD_CLEAN_0',
'/ssd/xulifeng/train_data/head_detection/fid/HeadVocFormat/FID_DID_HEAD_CLEAN_1',
'/ssd/xulifeng/train_data/head_detection/fid/HeadVocFormat/FID_DID_HEAD_CLEAN_2',
'/ssd/xulifeng/train_data/head_detection/fid/HeadVocFormat/HeadBoxDataFidChecked2']
'''
out_file = "/ssd/xulifeng/workspace/hd_densebox/train_data/nav_fullframe_lst.txt"
dataset_list = [
'/ssd/xieqiang/Data/HeadVocFormat/NavTrackingOffice_20190703',
'/ssd/xieqiang/Data/HeadVocFormat/NavTrackingAirport_20190704',
'/ssd/xieqiang/Data/HeadVocFormat/NavTrackingAirport_20190705',
'/ssd/xieqiang/Data/HeadVocFormat/NavTrackingAirport_20190706',
'/ssd/xieqiang/Data/HeadVocFormat/NavTrackingAirport_20190707',
'/ssd/xieqiang/Data/HeadVocFormat/NavTrackingAirport_20190708',
'/ssd/xieqiang/Data/HeadVocFormat/NavTrackingAirport_20190709',
'/ssd/xieqiang/Data/HeadVocFormat/NavTrackingAirport_20190710',
'/ssd/xieqiang/Data/HeadVocFormat/NavTrackingAirport_20190711',
'/ssd/xieqiang/Data/HeadVocFormat/NavTrackingAirport_20190712',
'/ssd/xieqiang/Data/HeadVocFormat/NavTrackingAirport_20190713',
'/ssd/xieqiang/Data/HeadVocFormat/NavTrackingAirport_20190714',
]
anno_type = 1 # fully labelled
fout = open(out_file, "wt")
for data_folder in dataset_list:
ori_anns_folder = os.path.join(data_folder, "Annotations")
ori_imgs_folder = os.path.join(data_folder, "JPEGImages")
imgs = list_all_files(ori_imgs_folder, exts=["jpg"])
for i, img_path in enumerate(imgs):
img_base_name = os.path.basename(img_path)
xml_base_name = os.path.splitext(img_base_name)[0] + ".xml"
ori_xml = os.path.join(ori_anns_folder, xml_base_name)
pascal_voc_ann = PascalVocAnn(xml=ori_xml)
boxes = pascal_voc_ann.get_boxes()
if len(boxes) == 0:
continue
vec = []
for b in boxes:
xmin, ymin, xmax, ymax = b[1:5]
w = xmax - xmin + 1
h = ymax - ymin + 1
sz = int(max(w, h) * 0.62)
x = int(xmin + (w - sz) * 0.5)
y = int(ymin + h - sz)
vec.extend([x, y, sz, sz])
line = img_path + "," + str(anno_type) + "," + ",".join(
map(str, vec)) + "\n"
fout.write(line)
fout.close()
def get_max_bbox_size(dataset_list):
res_max_size = 0
for data_folder in dataset_list:
ori_anns_folder = os.path.join(data_folder, "Annotations")
xmls = list_all_files(ori_anns_folder, exts=["xml"])
for i, xml_path in enumerate(xmls):
pascal_voc_ann = PascalVocAnn(xml=xml_path)
bboxes = pascal_voc_ann.get_boxes()
vec = []
for i, b in enumerate(bboxes):
xmin, ymin, xmax, ymax = b[1:5]
max_size = max(xmax - xmin, ymax - ymin)
if res_max_size < max_size:
res_max_size = max_size
return res_max_size
def draw_by_imgxml(imgpath, xmlpath, img_out):
_img_ph = imgpath
_xml_ph = xmlpath
print("imgpath:%s, xmlpath:%s" % (imgpath, xmlpath))
if not os.path.exists(_img_ph):
raise Exception("%s not exists" % (_img_ph))
if not os.path.exists(_xml_ph):
raise Exception("%s not exists" % (_xml_ph))
#img_ph = "example/originpic/ch00007_20190319_ch00007_20190319151500.mp4.cut.mp4_003000.jpg"
#xml_ph = "example/originpic/ch00007_20190319_ch00007_20190319151500.mp4.cut.mp4_003000.xml"
pascal_voc_ann = PascalVocAnn(xml=_xml_ph)
gtboxes = pascal_voc_ann.get_boxes()
print("basename:", os.path.basename(_img_ph))
print("splitext:", os.path.splitext(_img_ph))
image_basename = os.path.basename(_img_ph)
img = cv2.imread(_img_ph)
for bx in gtboxes:
img = cv2.rectangle(img, (int(bx[1]), int(bx[2])),
(int(bx[3]), int(bx[4])), (0, 0, 255), 1)
rect_img_savepath = img_out + "/" + os.path.splitext(
image_basename)[0] + "_gt.jpg"
cv2.imwrite(rect_img_savepath, img)
jpgs = list_all_files(imglist_prefix + "/JPEGImages/", exts=["jpg"])
label_dict = {}
dt_dict = {}
gtdict = {}
pred_dict = {}
save_dstpath = imglist_prefix + "/facessdITERATION_conf%s_iou%s_%s/" % (
input_conf, input_iouthd, save_timestamp)
if not os.path.exists(save_dstpath):
os.mkdir(save_dstpath)
for jpg_img in jpgs:
sample_basename = os.path.basename(jpg_img)
xml_temp = os.path.splitext(sample_basename)[0]
xml_sample_path = imglist_prefix + "/Annotations/" + xml_temp + ".xml"
pascal_voc_ann = PascalVocAnn(xml=xml_sample_path)
bboxes = pascal_voc_ann.get_boxes()
image = caffe.io.load_image(jpg_img)
image_basename = os.path.basename(jpg_img)
transformed_image = transformer.preprocess('data', image)
net.blobs['data'].data[...] = transformed_image
start = time.time()
# Forward pass.
detections = net.forward()['detection_out']
end = time.time()
logging.warning('Running time: {} Seconds'.format(end - start))
# Parse the outputs.
logging.info(detections)
det_label = detections[0, 0, :, 1]
def draw_bbox_on_image(nxmin,nymin,nxmax,nymax, img_filepath):
img = cv2.imread(img_filepath)
region_rect = cv2.rectangle(img, (int(nxmin),int(nymin)), (int(nxmax),int(nymax)), (0,0,255), 2)
rect_img_savepath = "example/gt_and_pred/"+os.path.basename(img_filepath)+"gt_and_pred.jpg"
cv2.imwrite(rect_img_savepath, region_rect)
if __name__ =='__main__':
parse = argparse.ArgumentParser()
parse.add_argument('--img_ph', type=str)
parse.add_argument('--xml_ph', type=str)
parse.add_argument('--env', type=str)
_args = parse.parse_args()
_img_ph= _args.img_ph;
_xml_ph= _args.xml_ph;
#img_ph = "example/originpic/ch00007_20190319_ch00007_20190319151500.mp4.cut.mp4_003000.jpg"
#xml_ph = "example/originpic/ch00007_20190319_ch00007_20190319151500.mp4.cut.mp4_003000.xml"
pascal_voc_ann = PascalVocAnn(xml=_xml_ph)
gtboxes = pascal_voc_ann.get_boxes()
print("basename:", os.path.basename(_img_ph))
print("splitext:", os.path.splitext(_img_ph))
image_basename = os.path.basename(_img_ph)
img = cv2.imread(_img_ph)
for bx in gtboxes:
img = cv2.rectangle(img, (int(bx[1]),int(bx[2])), (int(bx[3]),int(bx[4])), (0,0,255), 2)
rect_img_savepath = "./"+os.path.splitext(image_basename)[0]+"_gt.jpg"
cv2.imwrite(rect_img_savepath, img)
def crop_and_resize(img_path, ori_xml, newvocdir):
if not os.path.exists(newvocdir):
os.mkdir(newvocdir)
njpegpth = newvocdir + "/JPEGImages"
if not os.path.exists(njpegpth):
os.mkdir(njpegpth)
nannopth = newvocdir + "/Annotations"
if not os.path.exists(nannopth):
os.mkdir(nannopth)
img_filepath = img_path
lb_filepath = ori_xml
img = cv2.imread(img_filepath)
pascal_voc_ann = PascalVocAnn(xml=lb_filepath)
bboxes = pascal_voc_ann.get_boxes()
vec = []
for i, b in enumerate(bboxes):
print('reshape oribbox to retangle ...')
xmin, ymin, xmax, ymax = b[1:5]
w = xmax - xmin + 1
h = ymax - ymin + 1
sz = int(max(w, h) * 0.62)
x = int(xmin + (w - sz) * 0.5)
y = int(ymin + h - sz)
vec.extend([x, y, sz, sz])
xc = x + sz / 2
yc = y + sz / 2
print('crop img with 300x300 ...')
img_xmn = xc - IM_RUS if xc - IM_RUS >= 0 else 0
img_ymn = yc - IM_RUS if yc - IM_RUS >= 0 else 0
img_xmx = xc + IM_RUS if xc + IM_RUS < img.shape[
1] else img.shape[1] - 1
img_ymx = yc + IM_RUS if yc - IM_RUS < img.shape[
0] else img.shape[0] - 1
ioregion = img[img_ymn:img_ymx, img_xmn:img_xmx]
#print(img_filepath.split('.'))
#print(os.path.splitext(img_filepath))
new_anns_folder = os.path.join(newvocdir, "Annotations")
new_imgs_folder = os.path.join(newvocdir, "JPEGImages")
crop_img_name = os.path.splitext(
img_filepath)[0] + "_crop_%d.jpg" % (i)
crop_img_savepath = new_imgs_folder + "/" + os.path.basename(
crop_img_name)
cv2.imwrite(crop_img_savepath, ioregion)
print('rebase ori bbox pos ...')
#print("crop_img_savepath:%s", crop_img_savepath)
nxmin = x - img_xmn
nymin = y - img_ymn
nxmax = x + sz - 1 - img_xmn
nymax = y + sz - 1 - img_ymn
crop_xml_name = os.path.splitext(
os.path.basename(crop_img_savepath))[0] + ".xml"
crop_xml_savepath = new_anns_folder + "/" + crop_xml_name
newpascal_ann = PascalVocAnn(image_name=crop_img_savepath)
newpascal_ann.set_filename(file_name=crop_img_savepath)
newpascal_ann.set_size(size=[REQ_IMGSIZE, REQ_IMGSIZE, img.shape[2]])
newpascal_ann.add_object(object_class="head",
xmin=nxmin,
ymin=nymin,
xmax=nxmax,
ymax=nymax)
newpascal_ann.write_xml(crop_xml_savepath)
print('... done')
def crop_and_resize(img_path, ori_xml, newvocdir, gen_gt_rect=False):
if not os.path.exists(newvocdir):
os.mkdir(newvocdir)
njpegpth = newvocdir + "/JPEGImages"
if not os.path.exists(njpegpth):
os.mkdir(njpegpth)
nannopth = newvocdir + "/Annotations"
if not os.path.exists(nannopth):
os.mkdir(nannopth)
if gen_gt_rect:
rectpth = newvocdir + "/RECT_JPEGImages"
if not os.path.exists(rectpth):
os.mkdir(rectpth)
img_filepath = img_path
lb_filepath = ori_xml
img = cv2.imread(img_filepath)
pascal_voc_ann = PascalVocAnn(xml=lb_filepath)
bboxes = pascal_voc_ann.get_boxes()
h, w, c = pascal_voc_ann.get_size()
img_output_size = 72
rectangle_boxes = {}
rect_idx = 0
#print("raw bbox",bboxes)
for (i, b) in enumerate(bboxes):
xmin, ymin, xmax, ymax = b[1:5]
#print("raw xmin,ymin:(%d,%d), xmax,ymax:(%d,%d)"%(xmin,ymin,xmax,ymax))
w_raw = xmax - xmin + 1
h_raw = ymax - ymin + 1
sz = int(max(w_raw, h_raw) * 0.62)
x = int(xmin + (w_raw - sz) * 0.5)
y = int(ymin + h_raw - sz)
new_xmin = x
new_ymin = y
new_xmax = x + sz - 1
new_ymax = y + sz - 1
#print("x,y,sz:(%d,%d,%d)",x,y,sz)
new_xmin = new_xmin if new_xmin >= 0 else 0
new_ymin = new_ymin if new_ymin >= 0 else 0
new_xmax = new_xmax if new_xmax < w else w - 1
new_ymax = new_ymax if new_ymax < h else h - 1
#print("new xmin,ymin:(%d,%d), xmax,ymax:(%d,%d)"%(new_xmin,new_ymin,new_xmax,new_ymax))
rectangle_boxes[i] = [new_xmin, new_ymin, new_xmax, new_ymax, sz]
rect_idx = i
for (i, boxlist) in rectangle_boxes.items():
[xmin, ymin, xmax, ymax, sz] = boxlist
xc = xmin + sz / 2
yc = ymin + sz / 2
print('get crop info ...')
#print("xmin,ymin:(%d,%d), xmax,ymax:(%d,%d)"%(xmin,ymin,xmax,ymax))
rus = get_rus_by_head_box(box=[xc, yc, sz],
headrawsz=sz,
head_out_base_sz=20,
scale=0.2,
outputsize=72)
rus = int(rus)
rus = rus / 2
#print("xc:%d, yc:%d, xmin:%d, ymin:%d, rus:%d"%(xc,yc,xmin,ymin,rus))
img_xmn = xc - rus if xc - rus >= 0 else 0
img_ymn = yc - rus if yc - rus >= 0 else 0
img_xmx = xc + rus if xc + rus < img.shape[1] else img.shape[1] - 1
img_ymx = yc + rus if yc - rus < img.shape[0] else img.shape[0] - 1
#rectangle_boxes[rect_idx+1] = [img_xmn, img_ymn, img_xmx, img_ymx]
need_box = get_related_bboxes(rectangle_boxes,
[img_xmn, img_ymn, img_xmx, img_ymx])
ioregion_croped = img[img_ymn:img_ymx, img_xmn:img_xmx]
w_croped = ioregion_croped.shape[1]
h_croped = ioregion_croped.shape[0]
print('resize img from %sx%s to %sx%s ...' %
(rus, rus, img_output_size, img_output_size))
ioregion = cv2.resize(ioregion_croped,
(img_output_size, img_output_size),
interpolation=cv2.INTER_CUBIC)
print('... done')
new_anns_folder = os.path.join(newvocdir, "Annotations")
new_imgs_folder = os.path.join(newvocdir, "JPEGImages")
crop_img_name = os.path.splitext(
img_filepath)[0] + "_crop_%d.jpg" % (i)
crop_img_savepath = new_imgs_folder + "/" + os.path.basename(
crop_img_name)
cv2.imwrite(crop_img_savepath, ioregion)
print('rebase ori bbox pos ...')
rebased_need_box = []
#print("need_box",need_box)
for boxlist in need_box:
[box_xmin, box_ymin, box_xmax, box_ymax] = boxlist
reb_xmin = box_xmin - img_xmn
reb_ymin = box_ymin - img_ymn
reb_xmax = box_xmax - img_xmn
reb_ymax = box_ymax - img_ymn
reb_xmin = int(float(img_output_size) / w_croped * reb_xmin)
reb_ymin = int(float(img_output_size) / h_croped * reb_ymin)
reb_xmax = int(float(img_output_size) / w_croped * reb_xmax)
reb_ymax = int(float(img_output_size) / h_croped * reb_ymax)
rebased_need_box.append([reb_xmin, reb_ymin, reb_xmax, reb_ymax])
print('... done')
if gen_gt_rect:
print('draw rectangle ...')
region_rect = cv2.rectangle(ioregion, (int(nxmin), int(nymin)),
(int(nxmax), int(nymax)), (0, 255, 0),
2)
rect_imgs_folder = os.path.join(newvocdir, "RECT_JPEGImages")
rect_img_name = os.path.splitext(
img_filepath)[0] + "_rect_%d.jpg" % (i)
rect_img_savepath = rect_imgs_folder + "/" + os.path.basename(
rect_img_name)
cv2.imwrite(rect_img_savepath, region_rect)
crop_xml_name = os.path.splitext(
os.path.basename(crop_img_savepath))[0] + ".xml"
crop_xml_savepath = new_anns_folder + "/" + crop_xml_name
newpascal_ann = PascalVocAnn(image_name=crop_img_savepath)
newpascal_ann.set_filename(file_name=crop_img_savepath)
newpascal_ann.set_size(
size=[img_output_size, img_output_size, img.shape[2]])
for reb_box in rebased_need_box:
[reb_xmin, reb_ymin, reb_xmax, reb_ymax] = reb_box
newpascal_ann.add_object(object_class="head",
xmin=reb_xmin,
ymin=reb_ymin,
xmax=reb_xmax,
ymax=reb_ymax)
newpascal_ann.check_boxes()
newpascal_ann.write_xml(crop_xml_savepath)
print('... done')
def crop_and_resize(img_path, ori_xml, newvocdir, gen_gt_rect=False):
if not os.path.exists(newvocdir):
os.mkdir(newvocdir)
njpegpth = newvocdir + "/JPEGImages"
if not os.path.exists(njpegpth):
os.mkdir(njpegpth)
nannopth = newvocdir + "/Annotations"
if not os.path.exists(nannopth):
os.mkdir(nannopth)
if gen_gt_rect:
rectpth = newvocdir + "/RECT_JPEGImages"
if not os.path.exists(rectpth):
os.mkdir(rectpth)
img_filepath = img_path
lb_filepath = ori_xml
img = cv2.imread(img_filepath)
pascal_voc_ann = PascalVocAnn(xml=lb_filepath)
bboxes = pascal_voc_ann.get_boxes()
vec = []
print("Operating at ", img_filepath)
for i, b in enumerate(bboxes):
print('reshape oribbox to retangle ...')
xmin, ymin, xmax, ymax = b[1:5]
w = xmax - xmin + 1
h = ymax - ymin + 1
sz = int(max(w, h) * 0.62)
x = int(xmin + (w - sz) * 0.5)
y = int(ymin + h - sz)
vec.extend([x, y, sz, sz])
xc = x + sz / 2
yc = y + sz / 2
print('crop img with 300x300 ...')
print('input gtbox:', b)
img_xmn = xc - IM_RUS if xc - IM_RUS >= 0 else 0
img_ymn = yc - IM_RUS if yc - IM_RUS >= 0 else 0
img_xmx = xc + IM_RUS if xc + IM_RUS < img.shape[
1] else img.shape[1] - 1
img_ymx = yc + IM_RUS if yc - IM_RUS < img.shape[
0] else img.shape[0] - 1
ioregion_300 = img[img_ymn:img_ymx, img_xmn:img_xmx]
w300 = ioregion_300.shape[1]
h300 = ioregion_300.shape[0]
print('resize img from 300x300 to 128x128 ...')
ioregion = cv2.resize(ioregion_300, (128, 128),
interpolation=cv2.INTER_CUBIC)
print('... done')
#print(img_filepath.split('.'))
#print(os.path.splitext(img_filepath))
new_anns_folder = os.path.join(newvocdir, "Annotations")
new_imgs_folder = os.path.join(newvocdir, "JPEGImages")
crop_img_name = os.path.splitext(
img_filepath)[0] + "_crop_%d.jpg" % (i)
crop_img_savepath = new_imgs_folder + "/" + os.path.basename(
crop_img_name)
cv2.imwrite(crop_img_savepath, ioregion)
print('rebase ori bbox pos ...')
#print("crop_img_savepath:%s", crop_img_savepath)
nxmin = x - img_xmn
nymin = y - img_ymn
nxmax = x + sz - 1 - img_xmn
nymax = y + sz - 1 - img_ymn
print('... done')
print('resize bbox pos ...')
nxmin = int(128.0 / w300 * nxmin)
nymin = int(128.0 / h300 * nymin)
nxmax = int(128.0 / w300 * nxmax)
nymax = int(128.0 / h300 * nymax)
print('... done')
if gen_gt_rect:
print('draw rectangle ...')
region_rect = cv2.rectangle(ioregion, (int(nxmin), int(nymin)),
(int(nxmax), int(nymax)), (0, 255, 0),
2)
rect_imgs_folder = os.path.join(newvocdir, "RECT_JPEGImages")
rect_img_name = os.path.splitext(
img_filepath)[0] + "_rect_%d.jpg" % (i)
rect_img_savepath = rect_imgs_folder + "/" + os.path.basename(
rect_img_name)
cv2.imwrite(rect_img_savepath, region_rect)
crop_xml_name = os.path.splitext(
os.path.basename(crop_img_savepath))[0] + ".xml"
crop_xml_savepath = new_anns_folder + "/" + crop_xml_name
newpascal_ann = PascalVocAnn(image_name=crop_img_savepath)
newpascal_ann.set_filename(file_name=crop_img_savepath)
newpascal_ann.set_size(size=[128, 128, img.shape[2]])
newpascal_ann.add_object(object_class="head",
xmin=nxmin,
ymin=nymin,
xmax=nxmax,
ymax=nymax)
newpascal_ann.check_boxes()
newpascal_ann.write_xml(crop_xml_savepath)
print('... done')
def crop_and_resize(src_filename, dst_filename):
#img_filepath = 'ch01014_20190308_ch01014_20190308084110.mp4.cut.mp4_009000.jpg'
#lb_filepath = 'ch01014_20190308_ch01014_20190308084110.mp4.cut.mp4_009000.xml'
img_filepath = os.path.splitext(src_filename)[0] + '.jpg'
lb_filepath = os.path.splitext(src_filename)[0] + '.xml'
img = cv2.imread(img_filepath)
pascal_voc_ann = PascalVocAnn(xml=lb_filepath)
bboxes = pascal_voc_ann.get_boxes()
vec = []
for i, b in enumerate(bboxes):
xmin, ymin, xmax, ymax = b[1:5]
w = xmax - xmin + 1
h = ymax - ymin + 1
sz = int(max(w, h) * 0.62)
x = int(xmin + (w - sz) * 0.5)
y = int(ymin + h - sz)
vec.extend([x, y, sz, sz])
xc = x + sz / 2
yc = y + sz / 2
img_xmn = xc - IM_RUS if xc - IM_RUS >= 0 else 0
img_ymn = yc - IM_RUS if yc - IM_RUS >= 0 else 0
img_xmx = xc + IM_RUS if xc + IM_RUS < img.shape[
1] else img.shape[1] - 1
img_ymx = yc + IM_RUS if yc - IM_RUS < img.shape[
0] else img.shape[0] - 1
ioregion = img[img_ymn:img_ymx, img_xmn:img_xmx]
#print(img_filepath.split('.'))
#print(os.path.splitext(img_filepath))
crop_img_name = os.path.splitext(
img_filepath)[0] + "_crop_%d.jpg" % (i)
cv2.imwrite(crop_img_name, ioregion)
nxmin = x - img_xmn
nymin = y - img_ymn
nxmax = x + sz - 1 - img_xmn
nymax = y + sz - 1 - img_ymn
crop_xml_name = os.path.splitext(crop_img_name)[0] + ".xml"
newpascal_ann = PascalVocAnn(image_name=crop_img_name)
newpascal_ann.set_filename(file_name=crop_img_name)
newpascal_ann.set_size(size=[REQ_IMGSIZE, REQ_IMGSIZE, img.shape[2]])
newpascal_ann.add_object(object_class="head",
xmin=nxmin,
ymin=nymin,
xmax=nxmax,
ymax=nymax)
newpascal_ann.write_xml(crop_xml_name)
def gen_pascal_xml(image_path, size, box, xml_dst_path):
newpascal_ann = PascalVocAnn(image_name=image_path)
newpascal_ann.set_filename(file_name=image_path)
newpascal_ann.set_size(size=size)
nxmin, nymin, nxmax, nymax = box
newpascal_ann.add_object(object_class="head",
xmin=nxmin,
ymin=nymin,
xmax=nxmax,
ymax=nymax)
newpascal_ann.check_boxes()
newpascal_ann.write_xml(xml_dst_path)