def fov_match_and_fusion2(pred1,
pred2,
gt_label,
point1,
point2,
trust_first=True,
fusion_fuction=None,
args=None):
from detection_evaluation.nuscenes_eval_core import NuScenesEval
ret = []
pr1 = []
pr2 = []
gt = []
fov, _ = FOV(point1).caculate_iou(FOV(point2))
for ind, (frame_det1, frame_det2,
frame_gt) in enumerate(zip(pred1, pred2, gt_label)):
trust_frame = frame_det1 if trust_first else frame_det2
fov_det1 = np.empty((0, 8))
fov_det2 = np.empty((0, 8))
fov_gt = np.empty((0, 8))
for ind_un, car_point in enumerate(frame_det1.astype(np.float)):
#print(car_point)
if _polygon_contains_point(fov, (car_point[0], car_point[1])):
fov_det1 = np.vstack((fov_det1, car_point))
for ind_un, car_point in enumerate(frame_det2.astype(np.float)):
if _polygon_contains_point(fov, (car_point[0], car_point[1])):
fov_det2 = np.vstack((fov_det2, car_point))
for ind_un, car_point in enumerate(frame_gt.astype(np.float)):
if _polygon_contains_point(fov, (car_point[0], car_point[1])):
fov_gt = np.vstack((fov_gt, car_point))
matched, unmatched1, unmatched2 = match_pairs(
fov_det1.astype(np.float), fov_det2.astype(np.float))
#unmatched = np.vstack((unmatched1,unmatched2))
ret.append(np.vstack((matched, unmatched1, unmatched2)))
pr1.append(fov_det1)
pr2.append(fov_det2)
gt.append(fov_gt)
#delete_index = []
#print(ind,':',len(unmatched))
# for ind_un,car_point in enumerate(unmatched):
# if _polygon_contains_point(fov,(car_point[0],car_point[1])):
# if car_point.tolist() in trust_frame.tolist():
# continue
# else:
# delete_index.append(ind_un)
#tmp = np.delete(unmatched,obj=delete_index,axis=0)
#ret.append(np.vstack((matched,tmp)))
#ret.append(unmatched)
#print(len(tmp))
Eval = NuScenesEval('', '', args.format)
mAP_temp = Eval.my_evaluate(ret, gt)
Eval = NuScenesEval('', '', args.format)
mAP_1 = Eval.my_evaluate(pr1, gt)
Eval = NuScenesEval('', '', args.format)
mAP_2 = Eval.my_evaluate(pr2, gt)
return mAP_temp, mAP_1, mAP_2
def func_co(x):
x.sort()
Eval = NuScenesEval('', '', args.format)
fused_data = cu.matching_and_fusion(cam_test_list[x[0]],cam_test_list[x[1]])
fused_gt = cu.filt_gt_labels_tuple(cam_gt_list[x[0]],cam_gt_list[x[1]])
mAP_temp = Eval.my_evaluate(fused_data,fused_gt)
return 1- mAP_temp
def fuse_constellation(x):
#根据x的维度进行融合
x.sort()
size_n = x.shape[0]
main_cam = x[0]
gt_list = []
gt_list.append(cam_gt_list[main_cam])
for i in x[1:]:
fused_data = cu.matching_and_fusion(cam_test_list[main_cam],cam_test_list[i]) #融合
gt_list.append(cam_gt_list[i])
fused_gt = cu.filt_gt_labels_tuple(*gt_list)
Eval = NuScenesEval('', '', args.format)
mAP_temp = Eval.my_evaluate(fused_data,fused_gt)
return 1- mAP_temp
def fuse_constellation(x):
#根据x的维度进行融合
x.sort()
size_n = x.shape[0]
#fused_data = cam_test_list[x[0]]
#gt_list = []
#gt_list.append(cam_gt_list[main_cam])
new_cam_test = []
for i in x:
new_cam_test.append(cam_test_list[i])
fused_data = cu.matching_and_fusion_tuple(*new_cam_test, dbscan=dbscan)
# for i in x[1:]:
# fused_data = cu.matching_and_fusion(fused_data,cam_test_list[i]) #融合
#gt_list.append(cam_gt_list[i])
#fused_gt = cu.filt_gt_labels_tuple(*gt_list)
Eval = NuScenesEval('', '', args.format)
#print(fused_data == cam_test_list[x[0]])
mAP_temp = Eval.my_evaluate(fused_data, fused_gt)
return 1 - mAP_temp
# if mAP_temp > max_map:
# max_map = mAP_temp
# max_i,max_j = i,j
# print('temp max mAP: {}.......... time: ## i: {} j: {} k:{} z:{}'.format(max_map,i,j,k,z))
#print(mAP_temp)
max_map = 0
max_i, max_j = 0, 0
fused_gt = cu.filt_gt_labels_tuple(*cam_gt_list)
for i in range(34):
for j in range(i + 1, 34):
fused_data = cu.matching_and_fusion(cam_test_list[i],
cam_test_list[j]) #融合
Eval = NuScenesEval('', '', args.format)
#fused_gt = cu.filt_gt_labels_tuple(cam_gt_list[i],cam_gt_list[j],cam_gt_list[k])
#评估
mAP_temp = Eval.my_evaluate(fused_data, fused_gt)
if mAP_temp > max_map:
max_map = mAP_temp
max_i, max_j = i, j
print(
'temp max mAP: {}.......... time: ## i: {} j: {} '.
format(max_map, i, j))
#print(mAP_temp)
# max_map = 0
# max_i,max_j = 0,0
# fused_gt = cu.filt_gt_labels_tuple(*cam_gt_list)
# for i in range(34):
# for j in range(i+1,34):
# fused_data = cu.matching_and_fusion(cam_test_list[i],cam_test_list[j]) #融合
# for k in range(j+1,34):
# Eval = NuScenesEval('', '', args.format)
filt_start_time = time.time()
fused_gt = cu.filt_gt_labels_tuple(*cam_gt_list)
#ret = cu.filt_gt_labels(cam_gt_list[0],cam_gt_list[1])
#filt_time = time.time() - filt_start_time
#print("filt gt for 1 iter, time: ",filt_time)
filt_start_time = time.time()
dbscan = DBSCAN(eps = 1.8,min_samples=1)
#cam_test_index = [2,3,6,7,8,20,23,27,31]
cam_test_index = [ 0, 1, 2, 3, 4, 6, 7, 8, 9, 10, 12, 13, 14, 15, 16, 17, 19, 20, 21, 22, 23, 24, 25, 26,27,31,33]
new_cam_test = []
for i in cam_test_index:
new_cam_test.append(cam_test_list[i])
fused_test = cu.matching_and_fusion_tuple(*new_cam_test,dbscan=dbscan)
Eval = NuScenesEval('', '', args.format)
mAP_temp = Eval.my_evaluate(fused_test,fused_gt)
print(mAP_temp)
# X = cu.matching_and_fusion_tuple(*cam_test_list)
# data = np.stack([X[1][:,0],X[1][:,1]]).transpose()
# result = DBSCAN(eps = 1.6,min_samples=1).fit(data)
# y_pred = result.labels_ #每个元素的标签，同一聚类下的元素标签相同
# label_set = set(y_pred)
# filtered_preds = np.empty((0, 8))
# for lb in label_set:
# filter_n = np.asarray([lb])
# objs_in_cluster = X[1][np.in1d(y_pred, filter_n)]
# filtered_obj = cu.mean_fusion(objs_in_cluster)
# filtered_obj = cu.mean_fusion(objs_in_cluster)
# filtered_preds = np.vstack((filtered_preds, filtered_obj))
#print(X[0][:,0:2])
print(len(fused_gt[num_test]))
#print(y_pred)
#print("gt_num: ",len(fused_gt[1]))
#print(len(X[0]))
#print(X[0][:,0].shape)
#print(X[0].shape)
#print(data.shape)
cluster_num = len(label_set) - (1 if -1 in y_pred else 0)
print("cluster_num: ", cluster_num)
Eval = NuScenesEval('', '', args.format)
mAP_temp = Eval.my_evaluate([filtered_preds], [fused_gt[num_test]])
print('Evaluation: ', mAP_temp)
Eval = NuScenesEval('', '', args.format)
mAP_ori = Eval.my_evaluate([X[num_test]], [fused_gt[num_test]])
# fused_data = cu.matching_and_fusion(fused_data,cam_test_list[k])
# Eval = NuScenesEval('', '', args.format)
# mAP_temp = Eval.my_evaluate(fused_data[1],fused_gt[1])
# print('Evaluation2: ',mAP_temp)
plt.title('pred_num:{} cluster_num:{} mAP:{:.3f}'.format(
len(data[:, 0]), cluster_num, mAP_temp))
#plt.title('pred_num:{} gt_num:{} mAP:{:.3f}'.format(len(fused_data_ori[num_test][:,0]),len(fused_gt[num_test]),mAP_fused))
plt.scatter(data[:, 0], data[:, 1], marker='o', c=y_pred)
# plt.scatter(data[:,0], data[:,1], marker='o',c=data[:,1])
#plt.scatter(fused_data_ori[num_test][:,0], fused_data_ori[num_test][:,1], marker='o',c=fused_data_ori[num_test][:,1])
plt.show()