def _check_load_bbox(self, entry, instance_id):
"""
Check and load ground-truth labels
"""
entry_id = entry['id']
entry_id = [entry_id] if not isinstance(entry_id,
(list, tuple)) else entry_id
ann_ids = self.dataset.getAnnIds(imgIds=entry_id, iscrowd=None)
objs = self.dataset.loadAnns(ann_ids)
# check valid bboxes
valid_objs = []
width = entry['width']
height = entry['height']
_instance_count = 0
_redudant_count = 0
_amodal_count = 0
unique_bbs = set()
for obj in objs:
if obj.get('ignore', 0) == 1:
continue
if not self._use_crowd and obj.get('iscrowd', 0):
continue
if self._amodal:
xmin, ymin, xmax, ymax = bbox_xywh_to_xyxy(obj['bbox'])
if xmin < 0 or ymin < 0 or xmax > width or ymax > height:
_amodal_count += 1
else:
xmin, ymin, xmax, ymax = bbox_clip_xyxy(
bbox_xywh_to_xyxy(obj['bbox']), width, height)
if (xmin, ymin, xmax, ymax) in unique_bbs:
_redudant_count += 1
continue
box_w = (xmax - xmin)
box_h = (ymax - ymin)
area = box_w * box_h
if area <= self._min_object_area:
continue
# require non-zero box area
if xmax > xmin and ymax > ymin:
unique_bbs.add((xmin, ymin, xmax, ymax))
contiguous_cid = self.json_category_id_to_contiguous_id[
obj['category_id']]
valid_objs.append([
xmin, ymin, xmax, ymax, contiguous_cid,
instance_id + _instance_count
])
_instance_count += 1
if not valid_objs:
if not self._skip_empty:
# dummy invalid labels if no valid objects are found
valid_objs.append([-1, -1, -1, -1, -1, -1])
return valid_objs, _instance_count, _redudant_count, _amodal_count
def _loadtxt(self):
"""Load all annotations into memory."""
logging.debug("Loading %s annotations into memory...", str(self))
items = []
labels = []
for name in self._splits:
img_root = os.path.join(self._root, 'WIDER_{}',
'images').format(name)
anno_txt = os.path.join(self._root, 'wider_face_split',
'wider_face_{}_bbx_gt.txt').format(name)
with open(anno_txt, 'r') as f:
while True:
img_path = f.readline().strip()
if img_path == '':
break
num = int(f.readline().strip())
label = []
for i in range(num):
annos = [int(a) for a in f.readline().strip().split()]
annos[2] += 1 # inter-w to real-w
annos[3] += 1 # inter-h to real-h
if 'train' in self._splits:
attr = edict()
attr.width = annos[2]
attr.height = annos[3]
attr.blur = annos[4]
attr.expression = annos[5]
attr.illumination = annos[6]
attr.invalid = annos[7]
attr.occlusion = annos[8]
attr.pose = annos[9]
if self._ignore_face(attr):
continue
if len(annos) == 7:
# annos = bbox_xywh_to_xyxy(map(float, annos[:4])) + tuple(annos[4:])
annos = bbox_xywh_to_xyxy(
([float(item)
for item in annos[:4]])) + tuple(annos[4:])
else:
# annos = bbox_xywh_to_xyxy(map(float, annos[:4]))
annos = bbox_xywh_to_xyxy(
([float(item) for item in annos[:4]]))
label.append(annos)
if len(label) > 0 or not self._skip_empty:
items.append(os.path.join(img_root, img_path))
labels.append(np.array(label, dtype=np.float32))
return items, labels
def _check_load_bbox(self, coco, entry):
"""Check and load ground-truth labels"""
entry_id = entry['id']
# fix pycocotools _isArrayLike which don't work for str in python3
entry_id = [entry_id] if not isinstance(entry_id,
(list, tuple)) else entry_id
ann_ids = coco.getAnnIds(imgIds=entry_id, iscrowd=None)
objs = coco.loadAnns(ann_ids)
# check valid bboxes
valid_objs = []
width = entry['width']
height = entry['height']
for obj in objs:
if obj['area'] < self._min_object_area:
continue
if obj.get('ignore', 0) == 1:
continue
if not self._use_crowd and obj.get('iscrowd', 0):
continue
# convert from (x, y, w, h) to (xmin, ymin, xmax, ymax) and clip bound
xmin, ymin, xmax, ymax = bbox_clip_xyxy(
bbox_xywh_to_xyxy(obj['bbox']), width, height)
# require non-zero box area
if obj['area'] > 0 and xmax > xmin and ymax > ymin:
contiguous_cid = self.json_id_to_contiguous[obj['category_id']]
valid_objs.append([xmin, ymin, xmax, ymax, contiguous_cid])
if not valid_objs:
if self._allow_empty:
# dummy invalid labels if no valid objects are found
valid_objs.append([-1, -1, -1, -1, -1])
return valid_objs
def _load_txt(self):
"""load all annotations into memory"""
logging.debug('Loading %s annotations into memory....', str(self))
items = []
labels = []
data_root = os.path.join(self._root, self._dataset)
for name in self._splits:
anno_txt = os.path.join(data_root, 'yuncong_face_split',
"{}_{}.txt").format(self._dataset, name)
with open(anno_txt, 'r') as f:
while True:
record = f.readline().strip().split()
if len(record) == 0:
break
imgpath = record[0].strip()
num = int(record[1].strip())
label = []
for i in range(num):
bbox = record[2 + i * 5:2 + (i + 1) * 5]
bbox = [float(x) for x in bbox]
bbox = bbox_xywh_to_xyxy(bbox[1:])
label.append(bbox)
if len(label) > 0 or not self._skip_empty:
items.append(os.path.join(data_root, imgpath))
labels.append(np.array(label, dtype=np.float32))
return items, labels
def xywh_to_xyxy(self, pred):
_pred = []
for rec in pred:
rec[2] += 1 # inter-w to real-w
rec[3] += 1 # inter-h to real-h
rec = list(bbox_xywh_to_xyxy(rec[:4])) + [rec[4]]
_pred.append(rec)
return mx.nd.array(_pred, dtype='float64')
def parse_obj(obj):
contiguous_cid = self.json_id_to_contiguous[obj['category_id']]
if contiguous_cid >= self.num_class:
# not class of interest
return {"reason": 1}
if max(obj['keypoints']) == 0:
return {"reason": 2}
# convert from (x, y, w, h) to (xmin, ymin, xmax, ymax) and clip bound
xmin, ymin, xmax, ymax = bbox_clip_xyxy(
bbox_xywh_to_xyxy(obj['bbox']), width, height)
# require non-zero box area
if obj['area'] <= 0 or xmax <= xmin or ymax <= ymin:
return {"reason": 3}
# joints 3d: (num_joints, 3, 2); 3 is for x, y, z; 2 is for position, visibility
joints_3d = np.zeros((self.num_joints, 3, 2), dtype=np.float32)
for i in range(self.num_joints):
joints_3d[i, 0, 0] = obj['keypoints'][i * 3 + 0]
joints_3d[i, 1, 0] = obj['keypoints'][i * 3 + 1]
# joints_3d[i, 2, 0] = 0
visible = min(1, obj['keypoints'][i * 3 + 2])
joints_3d[i, :2, 1] = visible
# joints_3d[i, 2, 1] = 0
if np.sum(joints_3d[:, 0, 1]) < 1:
# no visible keypoint
return {"reason": 4}
if self._check_centers:
bbox_center, bbox_area = self._get_box_center_area(
(xmin, ymin, xmax, ymax))
kp_center, num_vis = self._get_keypoints_center_count(
joints_3d)
ks = np.exp(-2 *
np.sum(np.square(bbox_center - kp_center)) /
bbox_area)
if (num_vis / 80.0 + 47 / 80.0) > ks:
return {"reason": 5}
return {
'bbox': (xmin, ymin, xmax, ymax),
'joints_3d': joints_3d,
"reason": 0
}