def paser_func(img_id):
if not type(img_id) == type('123'):
img_id = img_id.numpy()
if type(img_id) == type(b'123'):
img_id = str(img_id, encoding='utf-8')
bboxs = id_bboxs_dict[img_id]
kps = id_kps_dict[img_id]
img = cv2.imread(os.path.join(img_path, img_id + '.jpg'))
# data aug
img, bboxs, kps = data_aug(img, bboxs, kps)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
# create center label
orih, oriw, oric = img.shape
neth, netw = params['height'], params['width']
outh, outw = neth // params['scale'], netw // params['scale']
centers, sigmas, whs = prepare_bbox(bboxs, orih, oriw, outh, outw)
keypoints, kps_sigmas = prepare_kps(kps, orih, oriw, outh, outw)
spm_label = SingleStageLabel(outh, outw, centers, sigmas, keypoints)
center_map, kps_map, kps_map_weight = spm_label()
# create img input
img = cv2.resize(img, (netw, neth), interpolation=cv2.INTER_CUBIC)
img = img.astype(np.float32) / 255.
return img, center_map, kps_map, kps_map_weight
def paser_func(img_id):
global id_bboxs_dict, params, img_path, id_kps_dict
if not type(img_id) == type('123'):
img_id = img_id.numpy()
if type(img_id) == type(b'123'):
img_id = str(img_id, encoding='utf-8')
bboxs = id_bboxs_dict[img_id]
kps = id_kps_dict[img_id]
img = cv2.imread(os.path.join(img_path, img_id+'.jpg'))
# data aug
# img, bboxs, kps = data_aug(img, bboxs, kps)
# padding img
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
h, w, c = img.shape
# 只在最右边或者最下边填充0, 这样不会影响box或者点的坐标值, 所以无需再对box或点的坐标做改变
if w > h:
img = cv2.copyMakeBorder(img, 0, w-h, 0, 0, cv2.BORDER_CONSTANT, value=(0, 0, 0))
else:
img = cv2.copyMakeBorder(img, 0, 0, 0, h-w, cv2.BORDER_CONSTANT, value=(0, 0, 0))
# create center label
orih, oriw, oric = img.shape
neth, netw = params['height'], params['width']
outh, outw = neth//params['scale'], netw//params['scale']
centers, sigmas, whs = prepare_bbox(bboxs, orih, oriw, outh, outw)
# center_encoder = centerEncoder(outh, outw, centers, sigmas, whs)
# center_label = center_encoder()
# create keypoints label
keypoints, kps_sigmas = prepare_kps(kps, orih, oriw, outh, outw)
# kps_encoder = kpsEncoder(outh, outw, centers, keypoints, sigmas=kps_sigmas)
# kps_label = kps_encoder()
spm_label = SingleStageLabel(outh, outw, centers, sigmas, keypoints)
label = spm_label()
# create img input
img = cv2.resize(img, (netw, neth), interpolation=cv2.INTER_CUBIC)
img = img.astype(np.float32) / 255. # conver to 0~1 tools if focal loss is right
# read by tensorflow, 不是很方便做iamge和box同时做数据增强
# image = tf.io.read_file(os.path.join(img_path, img_id+'.jpg'))
# image = tf.image.decode_jpeg(image)
# image = tf.image.convert_image_dtype(image, tf.float32)
# image = tf.image.resize_with_pad(image, netw, neth)
# label = np.concatenate([center_label, kps_label], axis=-1)
return img, label
def parser_func(img_id):
if type(img_id) != int:
img_id = int(img_id.numpy())
assert type(img_id) == int
img_info = coco.loadImgs(img_id)[0]
ann_ids = coco.getAnnIds(img_id, cat_ids)
annos = coco.loadAnns(ann_ids)
spm = SingleStageLabel(img_info, img_path, annos)
img, center_map, center_mask, kps_offset, kps_weight = spm(
params['height'], params['width'], params['scale'],
params['num_joints'])
return img, center_map, center_mask, kps_offset, kps_weight
# [x1, y1, w, h] -> [x1, y1, x2, y2]
bbox = [ann['bbox'][0], ann['bbox'][1], ann['bbox'][0]+ann['bbox'][2], ann['bbox'][1]+ann['bbox'][3]]
kps = ann['keypoints']
# print (kps)
assert len(kps) == 12 * 3
for i in range(12):
x = int(kps[i*3+0])
y = int(kps[i*3+1])
v = kps[i*3+2]
cv2.circle(img_ori, (x,y),4,colors[j%3],thickness=-1)
cv2.putText(img_ori, str(i), (x,y), cv2.FONT_HERSHEY_COMPLEX, 1,(0, 0, 250), 1)
cv2.imshow('ori', img_ori)
###########################################################
spm = SingleStageLabel(img_info, img_path, annos)
img, center_map, center_mask, kps_offset, kps_weight = spm(params['height'], params['width'], params['scale'], params['num_joints'])
cv2.imshow('center', center_map)
# data aug
# img, bboxs, kps = data_aug(img, bboxs, kps)
factor_x = 4
facotr_y = 4
spm_decoer = SpmDecoder(factor_x, facotr_y, 128, 128)
results = spm_decoer([center_map, kps_offset])
for j, result in enumerate(results):
center = result['center']
single_person_joints = result['joints']
cv2.circle(img, (int(center[0]), int(center[1])), 5, colors[j%3], thickness=-1)
0,
h - w,
cv2.BORDER_CONSTANT,
value=(0, 0, 0))
# create center label
orih, oriw, oric = img.shape
neth, netw = params['height'], params['width']
outh, outw = neth // params['scale'], netw // params['scale']
centers, sigmas, whs = prepare_bbox(bboxs, orih, oriw, outh, outw)
for center in centers:
print("ori center: ", center)
keypoints, kps_sigmas = prepare_kps(kps, orih, oriw, outh, outw)
spm_encoder = SingleStageLabel(outh, outw, centers, sigmas, keypoints)
spm_label = spm_encoder()
factor_x = netw / outw
facotr_y = neth / outh
spm_decoer = SpmDecoder(factor_x, facotr_y, outw, outh)
joints, decode_centers = spm_decoer(
[spm_label[..., 0:1], spm_label[..., 1:2 * 14 + 1]])
# create img input
img = cv2.resize(img, (netw, neth), interpolation=cv2.INTER_CUBIC)
for center in decode_centers:
print(center)
# x = int(center[0])
# y = int(center[1])