def plot_frame_bboxes(cam_name, frame_id):
gt_detection = reid_hash_map[cam_name]
yolo_detection = time_to_baseline_bbox(cam_name)
gt_rects = gt_detection[frame_id]
ab_rects, single_rects = [], []
for yolo_rect in yolo_detection[frame_id]:
max_iou = 0
for gt_rect in gt_rects:
max_iou = max(max_iou, general.IoU(gt_rect, yolo_rect))
if max_iou == 0:
single_rects.append(yolo_rect)
elif max_iou < 0.2:
ab_rects.append(yolo_rect)
if not (len(ab_rects) >= 1 and len(single_rects) >= 3): return
base_frame = general.get_frame(cam_name, frame_id - 1)
for rect in gt_rects:
left, top, width, height = rect
cv2.rectangle(base_frame, (left, top), (left + width, top + height),
(255, 0, 0), 3)
for rect in ab_rects:
left, top, width, height = [int(each) for each in rect]
cv2.rectangle(base_frame, (left, top), (left + width, top + height),
(0, 255, 0), 3)
for rect in single_rects:
left, top, width, height = [int(each) for each in rect]
cv2.rectangle(base_frame, (left, top), (left + width, top + height),
(0, 0, 255), 3)
print(frame_id)
cv2.imshow('image', base_frame)
cv2.waitKey()
cv2.destroyAllWindows()
def time_to_track_bbox_obj(cam_name):
result = {}
track_id_time_to_bbox = {}
for line in open(WORKSAPCE + DATA_PATH + cam_name + '/' + TRACK_PATH,
'r').readlines():
frame_id, obj_id, left, top, width, height, _, _, _, _ = [
round(float(each)) for each in line.split(',')
]
if obj_id not in track_id_time_to_bbox:
track_id_time_to_bbox[obj_id] = {
frame_id: (left, top, width, height)
}
else:
track_id_time_to_bbox[obj_id][frame_id] = (left, top, width,
height)
total_obj = len(track_id_time_to_bbox)
re_id_obj_count = 0
gt_time_to_bbox_obj = time_to_gt_bbox_obj(cam_name)
for obj_id in track_id_time_to_bbox:
reid_candidates = {}
for frame_id in track_id_time_to_bbox[obj_id]:
if frame_id not in gt_time_to_bbox_obj:
continue
for gt_bbox_obj in gt_time_to_bbox_obj[frame_id]:
track_rect = track_id_time_to_bbox[obj_id][frame_id]
gt_bbox, gt_obj = gt_bbox_obj[:4], gt_bbox_obj[-1]
IoU_score = general.IoU(track_rect, gt_bbox)
if IoU_score > 0.3:
if gt_obj not in reid_candidates:
reid_candidates[gt_obj] = IoU_score
else:
reid_candidates[gt_obj] += IoU_score
if len(reid_candidates) == 0:
new_obj_id = -1
else:
new_obj_id, obj_score = max(reid_candidates,
key=reid_candidates.get), max(
reid_candidates.keys())
if obj_score < 0.8:
new_obj_id = -1
else:
re_id_obj_count += 1
for frame_id in track_id_time_to_bbox[obj_id]:
bbox = track_id_time_to_bbox[obj_id][frame_id]
if frame_id not in result:
result[frame_id] = [bbox + (new_obj_id, )]
else:
result[frame_id].append(bbox + (new_obj_id, ))
print(cam_name, total_obj, re_id_obj_count)
return result
def assign_id(id_dict, target_dict):
time = target_dict.keys()
for t in time:
if t not in id_dict:
bboxes = target_dict[t]
id_bboxes = [each + (-1, ) for each in bboxes]
target_dict[t] = id_bboxes
continue
for idx, target_rect in enumerate(target_dict[t]):
best_fit_id, IoU_score = -1, 0
for gt_rect_id in id_dict[t]:
gt_rect, id = gt_rect_id[:4], gt_rect_id[-1]
IoU_tmp = general.IoU(target_rect, gt_rect)
if IoU_tmp < 0.1: continue
if IoU_tmp > IoU_score:
IoU_score = IoU_tmp
best_fit_id = id
target_dict[t][idx] = target_dict[t][idx] + (best_fit_id, )
return target_dict
def get_SVM_HashMap(gamma=2e-3, res_thres=2.0):
reid_hash_map = RegressionFilter.get_Regression_Hashmap(
res_thres=res_thres)
exist_max_oid = -1
for cam_name in reid_hash_map:
for _, obj in reid_hash_map[cam_name]:
exist_max_oid = max(exist_max_oid, obj)
# Logic to add some extra bbox from yolo detection
for cam_name in cameras:
time_to_bbox = {}
for frame_id, oid in reid_hash_map[cam_name]:
if frame_id not in time_to_bbox:
time_to_bbox[frame_id] = [
reid_hash_map[cam_name][(frame_id, oid)]
]
else:
time_to_bbox[frame_id].append(
reid_hash_map[cam_name][(frame_id, oid)])
for line in open(WORKSAPCE + DATA_PATH + cam_name + '/' +
DET_PATH).readlines():
frame_id, _, left, top, width, height, confidence, _, _, _ = [
float(each) for each in line.split(',')
]
if confidence < 0.05: continue
if left < 50 or top < 50 or left + width + 50 > cameras_shape[cam_name][1] or \
top + height + 50 > cameras_shape[cam_name][0]:
continue
frame_id, left, top, width, height = round(frame_id), round(
left), round(top), round(width), round(height)
add_this = True
if frame_id in time_to_bbox:
for exist_bbox in time_to_bbox[frame_id]:
if general.IoU(exist_bbox,
(left, top, width, height)) > 0.1:
add_this = False
break
if add_this:
exist_max_oid += 1
reid_hash_map[cam_name][(frame_id,
exist_max_oid)] = (left, top, width,
height)
outlier_dict = {cam: [] for cam in cameras}
unique_dict = {cam: [] for cam in cameras}
print("SVM function gamma input ", gamma)
for source_cam in cameras:
source_unique = set(reid_hash_map[source_cam].keys())
source_outlier = set()
for destination_cam in cameras:
if source_cam == destination_cam:
continue
svm_data, svm_label, data_id = prepare_svm_data(
reid_hash_map, source_cam, destination_cam, [0, 900])
false_pos = list(np.where(svm_label == 0)[0])
source_unique = source_unique.intersection(
set([data_id[i] for i in false_pos]))
clf = svm.SVC(kernel='rbf', class_weight='balanced',
gamma=gamma) # change to 2e-5 in overall setup
clf.fit(svm_data, svm_label)
y_pred = clf.predict(svm_data)
false_negative_pos = list(np.where((svm_label - y_pred) == -1)[0])
if len(false_negative_pos) == 0: continue
tmp_outlier = set([data_id[i] for i in false_negative_pos])
source_outlier = source_outlier.union(tmp_outlier)
outlier_dict[source_cam] = source_unique.intersection(source_outlier)
unique_dict[source_cam] = source_unique
outlier_num = 0
Multi_Hashmap = {}
for cam in cameras:
for key in outlier_dict[cam]:
del reid_hash_map[cam][key]
outlier_num += 1
Multi_Hashmap[cam] = reid_hash_map[cam]
print('SVM Outliers Number', outlier_num)
return Multi_Hashmap, reid_hash_map, outlier_dict, unique_dict
