def show_result(img, result, dataset='coco', score_thr=0.3, out_file=None):
img = mmcv.imread(img)
class_names = get_classes(dataset)
if isinstance(result, tuple):
bbox_result, segm_result = result
else:
bbox_result, segm_result = result, None
bboxes = np.vstack(bbox_result)
# draw segmentation masks
if segm_result is not None:
segms = mmcv.concat_list(segm_result)
inds = np.where(bboxes[:, -1] > score_thr)[0]
for i in inds:
color_mask = np.random.randint(0, 256, (1, 3), dtype=np.uint8)
mask = maskUtils.decode(segms[i]).astype(np.bool)
img[mask] = img[mask] * 0.5 + color_mask * 0.5
# draw bounding boxes
labels = [
np.full(bbox.shape[0], i, dtype=np.int32)
for i, bbox in enumerate(bbox_result)
]
labels = np.concatenate(labels)
mmcv.imshow_det_bboxes(img.copy(),
bboxes,
labels,
class_names=class_names,
score_thr=score_thr,
show=out_file is None)
def show_result(self,
data,
result,
img_norm_cfg,
dataset='coco',
score_thr=0.3):
if isinstance(result, tuple):
bbox_result, segm_result = result
else:
bbox_result, segm_result = result, None
img_tensor = data['img'][0]
img_metas = data['img_meta'][0].data[0]
imgs = tensor2imgs(img_tensor, **img_norm_cfg)
assert len(imgs) == len(img_metas)
if isinstance(dataset, str):
class_names = get_classes(dataset)
elif isinstance(dataset, (list, tuple)) or dataset is None:
class_names = dataset
else:
raise TypeError(
'dataset must be a valid dataset name or a sequence'
' of class names, not {}'.format(type(dataset)))
for img, img_meta in zip(imgs, img_metas):
h, w, _ = img_meta['img_shape']
img_show = img[:h, :w, :]
bboxes = np.vstack(bbox_result)
# draw segmentation masks
if segm_result is not None:
segms = mmcv.concat_list(segm_result)
inds = np.where(bboxes[:, -1] > score_thr)[0]
for i in inds:
color_mask = np.random.randint(
0, 256, (1, 3), dtype=np.uint8)
mask = maskUtils.decode(segms[i]).astype(np.bool)
img_show[mask] = img_show[mask] * 0.5 + color_mask * 0.5
# draw bounding boxes
labels = [
np.full(bbox.shape[0], i, dtype=np.int32)
for i, bbox in enumerate(bbox_result)
]
labels = np.concatenate(labels)
mmcv.imshow_det_bboxes(
img_show,
bboxes,
labels,
class_names=class_names,
score_thr=score_thr)
def _process_result(i, result):
# mask가 없을 때만 1번이다. mask 학습이었다면 2번이어야 한다... 지금은 귀찮음
bb_result, img_feats = result
class_names = get_classes(dataset)
labels = [
np.full(bbox.shape[0], i, dtype=np.int32)
for i, bbox in enumerate(bb_result)
]
labels = np.concatenate(labels)
bboxes = np.vstack(bb_result)
assert bboxes.ndim == 2
assert labels.ndim == 1
assert bboxes.shape[0] == labels.shape[0]
assert bboxes.shape[1] == 4 or bboxes.shape[1] == 5
assert bboxes.shape[1] == 5
scores = bboxes[:, -1]
bboxes = bboxes[:, :-1]
# print(i, scores, labels)
# if i == 23:
# exit()
_scores = list()
_bboxes = list()
_labels = list()
# CLASS = {"사람": 0, "소화기":1, "소화전":2, "차량":3, "자전거":4, "오토바이":5}
if i == 0:
for now_label_index, label in enumerate(labels):
limit_score = checked_score(label)
score = scores[now_label_index]
if score < limit_score:
continue
feature = img_feats[now_label_index]
bbox = bboxes[now_label_index]
# label, feature, bbox, score, id
object_info = ObjectInfo(label, feature, bbox, score,
objects_counter[label])
frame_dict[i][label].append(object_info)
objects_counter[label] += 1
_scores.append(score)
_labels.append(label)
_bboxes.append(bbox)
else:
for now_label_index, label in enumerate(labels):
# temp_dist = list()
limit_score = checked_score(label)
score = scores[now_label_index]
if score < limit_score:
continue
feature = img_feats[now_label_index]
bbox = bboxes[now_label_index]
_scores.append(score)
_labels.append(label)
_bboxes.append(bbox)
# 같은 라벨에 대해서 전부다 하는 것이 아닌 bbox가 가까운 아이들만 검사한다.
# 좌표 가져오기
x1, y1, x2, y2 = bbox
more_before_dist = list()
# 사람, 소화기, 소화전, 자전거 // 자동차,오토바이 따로 구분해서 처리
if label in [0, 1, 2, 4]:
# 원래는 검사하는 프레임을 i-20으로 20프레임 전만 검사하지만 현재는 테스트 중이므로 50 프레임을 검사한다.
# max(i - 20, 0)
for before_frame_index in range(max(i - 50, 0), i):
more_before_objects = frame_dict[before_frame_index]
more_before_label_objects = more_before_objects[label]
# debug 용
# if len(more_before_label_objects)>0:
# print("index:", before_frame_index, "label:", label)
for object_index, before_object in enumerate(
more_before_label_objects):
# 엇나갈 좌표의 범위는 사람 기준 dist_frame*50.0 px로 둔다.
# CLASS = {"사람": 0, "소화기":1, "소화전":2, "차량":3, "자전거":4, "오토바이":5}
# 소화기 소화전은 드론 속도만을 고려 하여 프레임당 30px 이동으로 준다.
dist_frame = i - before_frame_index
if label == 0:
dist_px = min(650.0, dist_frame * 200.0)
elif label == 1:
dist_px = min(650.0, dist_frame * 200.0)
elif label == 2:
dist_px = min(650.0, dist_frame * 200.0)
elif label == 4:
dist_px = min(650.0, dist_frame * 200.0)
b_x1, b_y1, b_x2, b_y2 = before_object.bbox
# a = x1 - b_x1 # 선 a의 길이
# b = y1 - b_y1 # 선 b의 길이
# print((x1, y1), "//", (b_x1, b_y1), "//", before_object.id)
# 테스트용 코드 작성
pred_coord_dist = 1000.0
predict_coordinate = before_object.get_predict_bbox(
)
if len(predict_coordinate) > 0:
center_x = (x2 + x1) / 2
center_y = (y2 + y1) / 2
_a = center_x - predict_coordinate[
0] # 선 a의 길이
_b = center_y - predict_coordinate[
1] # 선 b의 길이
# _a = x1 - predict_coordinate[0] # 선 a의 길이
# _b = y1 - predict_coordinate[1] # 선 b의 길이
pred_coord_dist = math.sqrt((_a * _a) +
(_b * _b))
if label == 0 or label == 1 or label == 2:
# 90 이었음
limit_coord_dist = 140
else:
limit_coord_dist = 200
if dist_frame != 1:
for num in range(1, dist_frame + 1):
limit_coord_dist += (limit_coord_dist *
np.exp(-num) * 2
) #np.exp2(-num))
edge_dist = 8
if x1 < edge_dist or y1 < edge_dist or \
x2 > img_size[0] - edge_dist or y2 > img_size[1] - edge_dist and \
(img_size[1] - 100) > pred_coord_dist: # 왜냐하면 왼쪽테두리로 나갔는데 오른쪽 테두리에 나왔을 때를 가정
limit_coord_dist *= 1.5
# bbox가 최대치 일때 비교와 이전 프레임의 feature 비교를 한다.
# 넓이는 최대 8*8 = 64 px 까지는 여유로 둔다.
max_dist_feat = 1000.0
if max((before_object.area - 64),
0) <= (x2 - x1) * (y2 - y1):
max_dist_feat = feature - before_object.max_bbox_feature
max_dist_feat = np.dot(max_dist_feat,
max_dist_feat)
_dist_feat = feature - before_object.feature
_dist_feat = np.dot(_dist_feat, _dist_feat)
dist_feat_list = [_dist_feat]
for _index, before_freature in enumerate(
before_object.feature_list):
_dist_feat = feature - before_freature
_dist_feat = np.dot(_dist_feat, _dist_feat)
# 임시로.. 오래된 이미지는 피쳐 신뢰도가 높을 수가 있으므로 약간 가중치를 준다.
_dist_feat = _dist_feat * np.exp(-_index)
dist_feat_list.append(_dist_feat)
dist_feat = min(dist_feat_list)
# 일단은 주위 100px 사이로만 검색 한다. (드론 이동이 있음) // 15 프레임까지는 거리 비교를 해본다.
# 또한 너무 비슷 하다면 기회를(?) 줘본다.
# print(dist_feat, max_dist_feat, dist_frame, "//", before_object.id, "//", pred_coord_dist, "//", b_x1, b_y1, "//", x1, y1)
# dist_px = dist_px if dist_frame < 15 else 100
# if min(x1+dist_px, img_size[0]) >= b_x1 and max(x1-dist_px,0) <= b_x1 and \
# min(y1 + dist_px, img_size[1]) >= b_y1 and max(y1 - dist_px, 0) <= b_y1 or \
# (pred_coord_dist < 130 and dist_frame < 30):
# 원래 13, 9 임
# if (dist_frame < 3 and min(x1 + limit_coord_dist, img_size[0]) >= b_x1 and max(x1 - limit_coord_dist, 0) <= b_x1 and \
# min(y1 + limit_coord_dist, img_size[1]) >= b_y1 and max(y1 - limit_coord_dist, 0) <= b_y1) or \
# pred_coord_dist < limit_coord_dist:
if (dist_frame < 2 and pred_coord_dist < dist_px
) or pred_coord_dist < limit_coord_dist:
# boundary 안에 있더라도 너무 안닮았으면 패스 한다. 200 이었음
if dist_feat < 150:
more_before_dist.append([
dist_feat * pred_coord_dist,
before_frame_index, object_index
])
elif dist_feat < 13:
more_before_dist.append([
dist_feat * pred_coord_dist,
before_frame_index, object_index
])
elif max_dist_feat < 8:
more_before_dist.append([
dist_feat * pred_coord_dist,
before_frame_index, object_index
])
else:
# 원본 max(i - 5, 0) 자동차 30프레임이였음
for before_frame_index in range(max(i - 100, 0), i):
# print("index:", before_frame_index)
more_before_objects = frame_dict[before_frame_index]
more_before_label_objects = more_before_objects[label]
for object_index, before_object in enumerate(
more_before_label_objects):
b_x1, b_y1 = before_object.bbox[:2]
# a = x1 - b_x1 # 선 a의 길이
# b = y1 - b_y1 # 선 b의 길이
# coord_dist = math.sqrt((a * a) + (b * b))
dist_frame = i - before_frame_index
# if label == 3:
# dist_px = min(650.0, dist_frame * 250.0)
# else:
# dist_px = min(650.0, dist_frame * 250.0)
# 테스트용 코드 작성
is_out_line = False
pred_coord_dist = 1500.0
predict_coordinate = before_object.get_predict_bbox(
)
if len(predict_coordinate) > 0:
pred_x = predict_coordinate[0]
pred_y = predict_coordinate[1]
center_x = (x2 + x1) / 2
center_y = (y2 + y1) / 2
_a = center_x - pred_x # 선 a의 길이
_b = center_y - pred_y # 선 b의 길이
# _a = x1 - predict_coordinate[0] # 선 a의 길이
# _b = y1 - predict_coordinate[1] # 선 b의 길이
pred_coord_dist = math.sqrt((_a * _a) +
(_b * _b))
# 차량 같은 경우 테두리를 아예 벗어나면 드론이 직접 돌리지 않는 이상 되돌아오지 않는다.
max_bbox_size = before_object.max_bbox_size
half_w = max_bbox_size[0] / 2
half_h = max_bbox_size[1] / 2
# pred 오차를 생각하여 100 px 이 벗어나면 다시 보이는 차들은 새로운 차임
line_limit = 100
if pred_x < 0 and pred_x + half_w <= 0 - line_limit:
is_out_line = True
elif pred_x > img_size[
0] and pred_x - half_w >= img_size[
0] + line_limit:
is_out_line = True
elif pred_y < 0 and pred_y + half_h <= 0 - line_limit:
is_out_line = True
elif pred_y > img_size[
1] and pred_y - half_h >= img_size[
1] + line_limit:
is_out_line = True
# 240 원래
limit_coord_dist = 230
if dist_frame != 1:
for num in range(1, dist_frame + 1):
limit_coord_dist += (limit_coord_dist *
np.exp2(-num))
limit_coord_dist = min(500.0, limit_coord_dist)
edge_dist = 8
if x1 < edge_dist or y1 < edge_dist or \
x2 > img_size[0] - edge_dist or y2 > img_size[1] - edge_dist and \
(img_size[1] - 100) > pred_coord_dist: # 왜냐하면 왼쪽테두리로 나갔는데 오른쪽 테두리에 나왔을 때를 가정:
limit_coord_dist *= 1.3
# bbox가 최대치 일때 비교와 이전 프레임의 feature 비교를 한다.
# 넓이는 최대 25*25 = 625 px 까지는 여유로 둔다.
max_dist_feat = 1000.0
if max((before_object.area - 625),
0) <= (x2 - x1) * (y2 - y1):
max_dist_feat = feature - before_object.max_bbox_feature
max_dist_feat = np.dot(max_dist_feat,
max_dist_feat)
# dist_feat = feature - before_object.feature
# dist_feat = np.dot(dist_feat, dist_feat)
_dist_feat = feature - before_object.feature
_dist_feat = np.dot(_dist_feat, _dist_feat)
dist_feat_list = [_dist_feat]
for _index, before_freature in enumerate(
before_object.feature_list):
_dist_feat = feature - before_freature
_dist_feat = np.dot(_dist_feat, _dist_feat)
# 임시로.. 오래된 이미지는 피쳐 신뢰도가 높을 수가 있으므로 약간 가중치를 준다.
_dist_feat = _dist_feat * np.exp(-_index)
dist_feat_list.append(_dist_feat)
dist_feat = min(dist_feat_list)
# 일단은 주위 100px 사이로만 검색 한다. (드론 이동이 있음) // 15 프레임까지는 거리 비교를 해본다.
# 또한 너무 비슷 하다면 기회를(?) 줘본다.
# print((x1, y1), "//", (b_x1, b_y1), "//", before_object.id, "//", dist_feat, max_dist_feat)
# print(dist_feat, max_dist_feat, dist_frame, "//", before_object.id, "//", pred_coord_dist)
# dist_px = dist_px if dist_frame < 10 else 300
# if min(x1 + dist_px, img_size[0]) >= b_x1 and max(x1 - dist_px, 0) <= b_x1 and \
# min(y1 + dist_px, img_size[1]) >= b_y1 and max(y1 - dist_px, 0) <= b_y1 or \
# (pred_coord_dist < limit_coord_dist and dist_frame < 30):
# 원래 feat 5, 3 이였음 지금 너무 피팅되서 숫자가 작게 나옴
if (dist_frame <= 4 and min(x1 + limit_coord_dist, img_size[0]) >= b_x1 and max(x1 - limit_coord_dist, 0) <= b_x1 and \
min(y1 + limit_coord_dist, img_size[1]) >= b_y1 and max(y1 - limit_coord_dist, 0) <= b_y1) or \
pred_coord_dist < limit_coord_dist and not (is_out_line and dist_frame > 5):
# boundary 안에 있더라도 너무 안닮았으면 패스 한다.
# 원래 180
if dist_feat < 170:
more_before_dist.append([
dist_feat * pred_coord_dist,
before_frame_index, object_index
])
elif dist_feat < 6:
more_before_dist.append([
dist_feat * pred_coord_dist,
before_frame_index, object_index
])
elif max_dist_feat < 4:
more_before_dist.append([
dist_feat * pred_coord_dist,
before_frame_index, object_index
])
if len(more_before_dist) == 0:
# print("??", i)
objects_counter[label] += 1
# label, feature, bbox, score
object_info = ObjectInfo(label, feature, bbox, score,
objects_counter[label])
frame_dict[i][label].append(object_info)
else:
more_before_dist = np.array(more_before_dist)
min_index = np.argmin(more_before_dist[:, 0])
choice_frame = int(more_before_dist[min_index][1])
choice_index = int(more_before_dist[min_index][2])
more_before_objects = frame_dict[choice_frame][label]
similarity_object = more_before_objects[choice_index]
similarity_object.update(feature, bbox, score)
frame_dict[i][label].append(similarity_object)
del more_before_objects[choice_index]
# 선택 받지 못한 예전 frame 들의 objects 들의 움직을 예측하여 업데이트 한다.
# 다만 30 프레임을 벗어난 아이들은 갱신하지 않는다.
for before_frame_index in range(max(i - 30, 0), i):
more_before_objects = frame_dict[before_frame_index]
more_before_label_objects = []
for label_index in range(6):
more_before_label_objects.extend(
more_before_objects[label_index])
for before_object in more_before_label_objects:
before_object.predict_bbox_update()
predictions['info'] = frame_dict[i]
predictions['bboxes'] = _bboxes
predictions['scores'] = _scores
predictions['labels'] = _labels
print(objects_counter)
if is_visualize:
visualize(os.path.join(dir_path, filename),
mmcv.imread(file_path),
predictions,
colors=colors,
mask_display=False,
class_names=class_names)
def _process_result(i, result):
# mask가 없을 때만 1번이다. mask 학습이었다면 2번이어야 한다... 지금은 귀찮음
bb_result, img_feats = result
class_names = get_classes(dataset)
labels = [
np.full(bbox.shape[0], i, dtype=np.int32)
for i, bbox in enumerate(bb_result)
]
labels = np.concatenate(labels)
bboxes = np.vstack(bb_result)
assert bboxes.ndim == 2
assert labels.ndim == 1
assert bboxes.shape[0] == labels.shape[0]
assert bboxes.shape[1] == 4 or bboxes.shape[1] == 5
assert bboxes.shape[1] == 5
scores = bboxes[:, -1]
bboxes = bboxes[:, :-1]
# print(i, scores, labels)
# if i == 23:
# exit()
_scores = list()
_bboxes = list()
_labels = list()
# CLASS = {"사람": 0, "소화기":1, "소화전":2, "차량":3, "자전거":4, "오토바이":5}
if i == 0:
for now_label_index, label in enumerate(labels):
limit_score = checked_score(label)
score = scores[now_label_index]
if score < limit_score:
continue
feature = img_feats[now_label_index]
bbox = bboxes[now_label_index]
# label, feature, bbox, score, id
object_info = ObjectInfo(label, feature, bbox, score,
objects_counter[label])
frame_dict[i][label].append(object_info)
objects_counter[label] += 1
_scores.append(score)
_labels.append(label)
_bboxes.append(bbox)
else:
for now_label_index, label in enumerate(labels):
# temp_dist = list()
limit_score = checked_score(label)
score = scores[now_label_index]
if score < limit_score:
continue
feature = img_feats[now_label_index]
bbox = bboxes[now_label_index]
_scores.append(score)
_labels.append(label)
_bboxes.append(bbox)
# 같은 라벨에 대해서 전부다 하는 것이 아닌 bbox가 가까운 아이들만 검사한다.
# 좌표 가져오기
x1, y1, x2, y2 = bbox
more_before_dist = list()
# 사람, 소화기, 소화전, 자전거 // 자동차,오토바이 따로 구분해서 처리
if label in [0, 1, 2, 4]:
# 원래는 검사하는 프레임을 i-20으로 20프레임 전만 검사하지만 현재는 테스트 중이므로 50 프레임을 검사한다.
# max(i - 20, 0)
for before_frame_index in range(max(i - 50, 0), i):
more_before_objects = frame_dict[before_frame_index]
more_before_label_objects = more_before_objects[label]
# debug 용
# if len(more_before_label_objects)>0:
# print("index:", before_frame_index, "label:", label)
for object_index, before_object in enumerate(
more_before_label_objects):
# 엇나갈 좌표의 범위는 사람 기준 dist_frame*50.0 px로 둔다.
# CLASS = {"사람": 0, "소화기":1, "소화전":2, "차량":3, "자전거":4, "오토바이":5}
# 소화기 소화전은 드론 속도만을 고려 하여 프레임당 30px 이동으로 준다.
dist_frame = i - before_frame_index
if label == 0:
dist_px = min(650.0, dist_frame * 120.0)
elif label == 1:
dist_px = min(650.0, dist_frame * 120.0)
elif label == 2:
dist_px = min(650.0, dist_frame * 120.0)
elif label == 3:
dist_px = min(650.0, dist_frame * 200.0)
elif label == 4:
dist_px = min(650.0, dist_frame * 120.0)
else:
dist_px = min(650.0, dist_frame * 200.0)
b_x1, b_y1, b_x2, b_y2 = before_object.bbox
# a = x1 - b_x1 # 선 a의 길이
# b = y1 - b_y1 # 선 b의 길이
# print((x1, y1), "//", (b_x1, b_y1), "//", before_object.id)
# 테스트용 코드 작성
pred_coord_dist = 1000.0
predict_coordinate = before_object.get_predict_bbox(
)
if len(predict_coordinate) > 0:
_a = x1 - predict_coordinate[0] # 선 a의 길이
_b = y1 - predict_coordinate[1] # 선 b의 길이
pred_coord_dist = math.sqrt((_a * _a) +
(_b * _b))
# bbox가 최대치 일때 비교와 이전 프레임의 feature 비교를 한다.
# 넓이는 최대 8*8 = 64 px 까지는 여유로 둔다.
max_dist_feat = 1000.0
if max((before_object.area - 64),
0) <= (x2 - x1) * (y2 - y1):
max_dist_feat = feature - before_object.max_bbox_feature
max_dist_feat = np.dot(max_dist_feat,
max_dist_feat)
_dist_feat = feature - before_object.feature
_dist_feat = np.dot(_dist_feat, _dist_feat)
dist_feat_list = [_dist_feat]
for before_freature in before_object.feature_list:
_dist_feat = feature - before_freature
_dist_feat = np.dot(_dist_feat, _dist_feat)
dist_feat_list.append(_dist_feat)
dist_feat = min(dist_feat_list)
# 일단은 주위 100px 사이로만 검색 한다. (드론 이동이 있음) // 15 프레임까지는 거리 비교를 해본다.
# 또한 너무 비슷 하다면 기회를(?) 줘본다.
# print(dist_feat, max_dist_feat, dist_frame, "//", before_object.id, "//", pred_coord_dist)
dist_px = dist_px if dist_frame < 15 else 100
if min(x1+dist_px,1920) >= b_x1 and max(x1-dist_px,0) <= b_x1 and \
min(y1 + dist_px, 1080) >= b_y1 and max(y1 - dist_px, 0) <= b_y1 or \
(pred_coord_dist < 150 and dist_frame < 30):
# boundary 안에 있더라도 너무 안닮았으면 패스 한다.
if dist_feat < 200:
more_before_dist.append([
dist_feat * pred_coord_dist,
before_frame_index, object_index
])
elif dist_feat < 12:
more_before_dist.append([
dist_feat * pred_coord_dist,
before_frame_index, object_index
])
elif max_dist_feat < 8:
more_before_dist.append([
dist_feat * pred_coord_dist,
before_frame_index, object_index
])
else:
# 원본 max(i - 5, 0) 자동차 15프레임이였음 (10프레임이 현재는 가장 스코어가 좋음)
for before_frame_index in range(max(i - 30, 0), i):
# print("index:", before_frame_index)
more_before_objects = frame_dict[before_frame_index]
more_before_label_objects = more_before_objects[label]
for object_index, before_object in enumerate(
more_before_label_objects):
b_x1, b_y1 = before_object.bbox[:2]
# a = x1 - b_x1 # 선 a의 길이
# b = y1 - b_y1 # 선 b의 길이
# coord_dist = math.sqrt((a * a) + (b * b))
dist_frame = i - before_frame_index
if label == 3:
dist_px = min(650.0, dist_frame * 210.0)
else:
dist_px = min(650.0, dist_frame * 210.0)
# 테스트용 코드 작성
pred_coord_dist = 1000.0
predict_coordinate = before_object.get_predict_bbox(
)
if len(predict_coordinate) > 0:
_a = x1 - predict_coordinate[0] # 선 a의 길이
_b = y1 - predict_coordinate[1] # 선 b의 길이
pred_coord_dist = math.sqrt((_a * _a) +
(_b * _b))
# bbox가 최대치 일때 비교와 이전 프레임의 feature 비교를 한다.
# 넓이는 최대 25*25 = 625 px 까지는 여유로 둔다.
max_dist_feat = 1000.0
if max((before_object.area - 625),
0) <= (x2 - x1) * (y2 - y1):
max_dist_feat = feature - before_object.max_bbox_feature
max_dist_feat = np.dot(max_dist_feat,
max_dist_feat)
# dist_feat = feature - before_object.feature
# dist_feat = np.dot(dist_feat, dist_feat)
_dist_feat = feature - before_object.feature
_dist_feat = np.dot(_dist_feat, _dist_feat)
dist_feat_list = [_dist_feat]
for before_freature in before_object.feature_list:
_dist_feat = feature - before_freature
_dist_feat = np.dot(_dist_feat, _dist_feat)
dist_feat_list.append(_dist_feat)
dist_feat = min(dist_feat_list)
# 일단은 주위 100px 사이로만 검색 한다. (드론 이동이 있음) // 15 프레임까지는 거리 비교를 해본다.
# 또한 너무 비슷 하다면 기회를(?) 줘본다.
# print((x1, y1), "//", (b_x1, b_y1), "//", before_object.id, "//", dist_feat, max_dist_feat)
# print(dist_feat, max_dist_feat, dist_frame, "//", before_object.id, "//", pred_coord_dist)
dist_px = dist_px if dist_frame < 10 else 300
if min(x1 + dist_px, 1920) >= b_x1 and max(x1 - dist_px, 0) <= b_x1 and \
min(y1 + dist_px, 1080) >= b_y1 and max(y1 - dist_px, 0) <= b_y1 or \
(pred_coord_dist < 300 and dist_frame < 30):
# boundary 안에 있더라도 너무 안닮았으면 패스 한다.
if dist_feat < 180:
more_before_dist.append([
dist_feat * pred_coord_dist,
before_frame_index, object_index
])
elif dist_feat < 6:
more_before_dist.append([
dist_feat * pred_coord_dist,
before_frame_index, object_index
])
elif max_dist_feat < 3:
more_before_dist.append([
dist_feat * pred_coord_dist,
before_frame_index, object_index
])
if len(more_before_dist) == 0:
# print("??", i)
objects_counter[label] += 1
# label, feature, bbox, score
object_info = ObjectInfo(label, feature, bbox, score,
objects_counter[label])
frame_dict[i][label].append(object_info)
else:
more_before_dist = np.array(more_before_dist)
min_index = np.argmin(more_before_dist[:, 0])
choice_frame = int(more_before_dist[min_index][1])
choice_index = int(more_before_dist[min_index][2])
more_before_objects = frame_dict[choice_frame][label]
similarity_object = more_before_objects[choice_index]
similarity_object.update(feature, bbox, score)
frame_dict[i][label].append(similarity_object)
del more_before_objects[choice_index]
# 선택 받지 못한 예전 frame 들의 objects 들의 움직을 예측하여 업데이트 한다.
# 다만 30 프레임을 벗어난 아이들은 갱신하지 않는다.
for before_frame_index in range(max(i - 30, 0), i):
more_before_objects = frame_dict[before_frame_index]
more_before_label_objects = []
for label_index in range(6):
more_before_label_objects.extend(
more_before_objects[label_index])
for before_object in more_before_label_objects:
before_object.predict_bbox_update()
predictions['info'] = frame_dict[i]
predictions['bboxes'] = _bboxes
predictions['scores'] = _scores
predictions['labels'] = _labels
print(objects_counter)
if is_visualize:
visualize(os.path.join(dir_path, filename),
mmcv.imread(file_path),
predictions,
colors=colors,
mask_display=False,
class_names=class_names)
dataset = 'kmu'
score_thr = 0.5
predictions = dict()
for i, result in enumerate(
inference_detector(model, imgs, cfg, device='cuda:0')):
print(i, imgs[i])
# print(result)
# mask 제외 시킨다 지금은
result = result[0]
file_path = imgs[i]
filename = os.path.split(file_path)[-1]
class_names = get_classes(dataset)
labels = [
np.full(bbox.shape[0], i, dtype=np.int32)
for i, bbox in enumerate(result)
]
labels = np.concatenate(labels)
bboxes = np.vstack(result)
img = mmcv.imread(file_path)
# if filename in img_info:
# gt_bboxes = img_info[filename]
#
# for gt_bbox in gt_bboxes:
# img = cv2.rectangle(img, tuple(gt_bbox[:2]), tuple(gt_bbox[2:]), (255,0,0), 2)
assert bboxes.ndim == 2