def _parse_function(imgId, is_train, ann):
global TRAIN_ANNO
global VALID_ANNO
anno = ann
img_meta = anno.loadImgs([imgId])[0]
anno_ids = anno.getAnnIds(imgIds=imgId)
img_anno = anno.loadAnns(anno_ids)
idx = img_meta['id']
img_path = join(IMAGE_PATH, img_meta['file_name'])
img_meta_data = CocoMetadata(idx, img_path, img_meta, img_anno, sigma=6.0)
display_original_image(img_path, img_anno)
heatmap = img_meta_data.get_heatmap(
(img_meta_data.height, img_meta_data.width))
leng = len(img_meta_data.joint_list[0])
joint = img_meta_data.joint_list[0]
plt.figure(1, figsize=(30, 30))
n_columns = 5
n_rows = math.ceil(leng / n_columns)
b, g, r = cv2.split(img_meta_data.img)
img_resize = cv2.merge((r, g, b))
i = 5
for i in range(leng):
# img_resize = copy.copy(img_meta_data.img)
if i > 4:
#plt.subplot(n_rows, n_columns, i + 1)
#plt.title('Joint Point' + str(i))
img_resize = cv2.circle(img_resize, joint[i], 5, (0, 255, 0),
cv2.FILLED) #CV2.filled #채워진 점
plt.imshow(img_resize)
#추가 coco api사용해서 점을 잇는 함수
anno.showAnns(img_anno)
set_p = [point for point in img_meta_data.joint_list[0]]
head_point = cv2.circle(
img_resize, ((set_p[2][0] + set_p[1][0]) // 2,
(set_p[2][1] + set_p[1][1]) * 3 // 2 - set_p[0][1] * 2),
5, (255, 0, 0), cv2.FILLED)
neck_point = cv2.circle(
img_resize, ((set_p[6][0] + set_p[5][0]) // 2,
(set_p[6][1] + set_p[5][1]) // 3 + set_p[0][1] // 3), 5,
(255, 0, 0), cv2.FILLED)
plt.imshow(head_point)
plt.imshow(neck_point)
plt.show()
def _parse_function(imgId, is_train, ann=None):
"""
:param imgId:
:return:
"""
global TRAIN_ANNO
global VALID_ANNO
if ann is not None:
if is_train == True:
TRAIN_ANNO = ann
else:
VALID_ANNO = ann
else:
if is_train == True:
anno = TRAIN_ANNO
else:
anno = VALID_ANNO
img_meta = anno.loadImgs([imgId])[0]
anno_ids = anno.getAnnIds(imgIds=imgId)
img_anno = anno.loadAnns(anno_ids)
idx = img_meta['id']
img_path = join(BASE, img_meta['file_name'])
img_meta_data = CocoMetadata(idx, img_path, img_meta, img_anno, sigma=6.0)
img_meta_data = pose_random_scale(img_meta_data)
img_meta_data = pose_rotation(img_meta_data)
img_meta_data = pose_flip(img_meta_data)
img_meta_data = pose_resize_shortestedge_random(img_meta_data)
img_meta_data = pose_crop_random(img_meta_data)
return pose_to_img(img_meta_data)
def _parse_function(self, imgId, ann=None):
"""
:param imgId:
:return:
"""
global TRAIN_ANNO
if ann is not None:
TRAIN_ANNO = ann
# print('imgId = %s' % imgId)
img_meta = TRAIN_ANNO.loadImgs([imgId])[0]
anno_ids = TRAIN_ANNO.getAnnIds(imgIds=imgId)
img_anno = TRAIN_ANNO.loadAnns(anno_ids)
idx = img_meta['id']
filename_item_list = img_meta['file_name'].split('/')
filename = filename_item_list[1] + '/' + filename_item_list[2]
img_path = join(DATASET_DIR, filename)
img_meta_data = CocoMetadata(idx=idx,
img_path=img_path,
img_meta=img_meta,
annotations=img_anno,
sigma=preproc_config.heatmap_std)
# print('joint_list = %s' % img_meta_data.joint_list)
images, labels = self.image_preprocessing_fn(
img_meta_data=img_meta_data, preproc_config=preproc_config)
return images, labels
def _parse_function(imgId, is_train, ann=None):
"""
:param imgId:
:return:
"""
global TRAIN_ANNO
global VALID_ANNO
if ann is not None:
if is_train == True:
TRAIN_ANNO = ann
else:
VALID_ANNO = ann
else:
if is_train == True:
anno = TRAIN_ANNO
else:
anno = VALID_ANNO
img_meta = anno.loadImgs([imgId])[0]
anno_ids = anno.getAnnIds(imgIds=imgId)
img_anno = anno.loadAnns(anno_ids)
idx = img_meta['id']
img_path = join(BASE, img_meta['file_name'])
if SEG_MASK:
mask_path = join(BASE_MASK_PATH, img_meta['file_name'])
else:
mask_path = None
# print('img_path==>', is_train, img_path)
# exit()
img_meta_data = CocoMetadata(idx, img_path, img_meta, img_anno, sigma=6.0, mask_path=mask_path) # 加载图像和kp标注
# 先数据增强，再把kp标注转换为heatmap
img_meta_data = pose_random_scale(img_meta_data, SEG_MASK)
img_meta_data = pose_rotation(img_meta_data, SEG_MASK)
img_meta_data = pose_flip(img_meta_data, SEG_MASK) #todo:multi task1.22改到这里
img_meta_data = pose_resize_shortestedge_random(img_meta_data, SEG_MASK)
img_meta_data = pose_crop_random(img_meta_data, SEG_MASK)
return pose_to_img(img_meta_data, imgId)
def _parse_function(self, imgId, ann=None):
"""
:param imgId:
:return:
"""
global TRAIN_ANNO
if ann is not None:
TRAIN_ANNO = ann
img_meta = TRAIN_ANNO.loadImgs([imgId])[0]
anno_ids = TRAIN_ANNO.getAnnIds(imgIds=imgId)
img_anno = TRAIN_ANNO.loadAnns(anno_ids)
idx = img_meta['id']
filename_item_list = img_meta['file_name'].split('/')
filename = filename_item_list[1] + '/' + filename_item_list[2]
if self.is_testcode:
# for test_data_loader_coco.py -----------------------
img_path = join(self.data_dir, filename)
else:
# for actual training -----------------------
img_path = join(FLAGS.data_dir, filename)
img_meta_data = CocoMetadata(idx=idx,
img_path=img_path,
img_meta=img_meta,
annotations=img_anno,
sigma=preproc_config.heatmap_std)
# print('joint_list = %s' % img_meta_data.joint_list)
images, labels = self.image_preprocessing_fn(
img_meta_data=img_meta_data,
preproc_config=preproc_config,
is_training=self.is_training)
return images, labels