-
Notifications
You must be signed in to change notification settings - Fork 0
/
eval_recall.py
executable file
·234 lines (195 loc) · 10.1 KB
/
eval_recall.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
import os
import torch
from net import RINet_attention_cir_pad, RINet_attention_cons_pad
from database import seq2pc_test_recall
import numpy as np
from torch.utils.data import DataLoader
from tqdm import tqdm
from sklearn import metrics
from matplotlib import pyplot as plt
import sys
import time
import argparse
from MAE import *
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
def recall(seq='00', model_file=None,model_mae_file=None,
velo_desc_folder=None, img_desc_folder=None,
pose_file="./data/pose_kitti/00.txt"):
img_desc_folder_0=os.path.join(img_desc_folder,'0')
img_desc_folder_5=os.path.join(img_desc_folder,'5')
img_desc_folder_10=os.path.join(img_desc_folder,'10')
img_desc_folder_15=os.path.join(img_desc_folder,'15')
img_desc_folder_cb=os.path.join(img_desc_folder,'combine')
velo_desc_folder_0=os.path.join(velo_desc_folder,'0')
velo_desc_folder_1=os.path.join(velo_desc_folder,'1')
velo_desc_folder_2=os.path.join(velo_desc_folder,'2')
velo_desc_folder_3=os.path.join(velo_desc_folder,'3')
velo_desc_folder_4=os.path.join(velo_desc_folder,'4')
velo_desc_folder_5=os.path.join(velo_desc_folder,'5')
velo_desc_folder_6=os.path.join(velo_desc_folder,'6')
velo_desc_folder_7=os.path.join(velo_desc_folder,'7')
if seq=='08':
net=RINet_attention_cir_pad()
else:
net = RINet_attention_cons_pad()
net.load(model_file)
net.to(device=device)
net.eval()
mask_ratio = 0.1
model_mae = MAE_ViT(mask_ratio = mask_ratio).to(device)
checkpoint = torch.load(model_mae_file)
model_mae.load_state_dict(checkpoint['state_dict'])
model_mae.eval()
img_desc_file_0=os.path.join(img_desc_folder_0, seq+'.bin')
img_desc_file_5=os.path.join(img_desc_folder_5, seq+'.bin')
img_desc_file_10=os.path.join(img_desc_folder_10, seq+'.bin')
img_desc_file_15=os.path.join(img_desc_folder_15, seq+'.bin')
img_desc_file_cb=os.path.join(img_desc_folder_cb, seq+'.bin')
velo_desc_file_0=os.path.join(velo_desc_folder_0, seq+'.bin')
velo_desc_file_1=os.path.join(velo_desc_folder_1, seq+'.bin')
velo_desc_file_2=os.path.join(velo_desc_folder_2, seq+'.bin')
velo_desc_file_3=os.path.join(velo_desc_folder_3, seq+'.bin')
velo_desc_file_4=os.path.join(velo_desc_folder_4, seq+'.bin')
velo_desc_file_5=os.path.join(velo_desc_folder_5, seq+'.bin')
velo_desc_file_6=os.path.join(velo_desc_folder_6, seq+'.bin')
velo_desc_file_7=os.path.join(velo_desc_folder_7, seq+'.bin')
img_desc_0=np.fromfile(img_desc_file_0, dtype=np.float32).reshape(-1,12,360)#img_desc_0,img_desc_5,img_desc_10,img_desc_15的长度是一样的
img_desc_5=np.fromfile(img_desc_file_5, dtype=np.float32).reshape(-1,12,360)
img_desc_10=np.fromfile(img_desc_file_10, dtype=np.float32).reshape(-1,12,360)
img_desc_15=np.fromfile(img_desc_file_15, dtype=np.float32).reshape(-1,12,360)
img_desc_cb=np.fromfile(img_desc_file_cb, dtype=np.float32).reshape(-1,12,360)
velo_desc_0=np.fromfile(velo_desc_file_0, dtype=np.float32).reshape(-1,12,360)
velo_desc_1=np.fromfile(velo_desc_file_1, dtype=np.float32).reshape(-1,12,360)
velo_desc_2=np.fromfile(velo_desc_file_2, dtype=np.float32).reshape(-1,12,360)
velo_desc_3=np.fromfile(velo_desc_file_3, dtype=np.float32).reshape(-1,12,360)
velo_desc_4=np.fromfile(velo_desc_file_4, dtype=np.float32).reshape(-1,12,360)
velo_desc_5=np.fromfile(velo_desc_file_5, dtype=np.float32).reshape(-1,12,360)
velo_desc_6=np.fromfile(velo_desc_file_6, dtype=np.float32).reshape(-1,12,360)
velo_desc_7=np.fromfile(velo_desc_file_7, dtype=np.float32).reshape(-1,12,360)
#转化为tensor 上传到GPU中，喂给dataset再进一步喂给dataloader
img_desc_0=torch.from_numpy(img_desc_0).to(device)
img_desc_5=torch.from_numpy(img_desc_5).to(device)
img_desc_10=torch.from_numpy(img_desc_10).to(device)
img_desc_15=torch.from_numpy(img_desc_15).to(device)
img_desc_cb=torch.from_numpy(img_desc_cb).to(device)
velo_desc_0=torch.from_numpy(velo_desc_0).to(device)
velo_desc_1=torch.from_numpy(velo_desc_1).to(device)
velo_desc_2=torch.from_numpy(velo_desc_2).to(device)
velo_desc_3=torch.from_numpy(velo_desc_3).to(device)
velo_desc_4=torch.from_numpy(velo_desc_4).to(device)
velo_desc_5=torch.from_numpy(velo_desc_5).to(device)
velo_desc_6=torch.from_numpy(velo_desc_6).to(device)
velo_desc_7=torch.from_numpy(velo_desc_7).to(device)
poses = np.genfromtxt(pose_file)
poses = poses[:, [3, 11]]
inner = 2*np.matmul(poses, poses.T)
xx = np.sum(poses**2, 1, keepdims=True)
dis = xx-inner+xx.T
dis = np.sqrt(np.abs(dis))
id_pos = np.argwhere(dis <= 5)
id_pos = id_pos[id_pos[:, 0]-id_pos[:, 1] > 50]
pos_dict = {}
#删除一些假回环，由车辆停滞时间过长导致的
for v in id_pos: #id_pos的最小值是0
for ii in range(v[0]-v[1]):
if dis[v[0],v[0]-ii-1]>5:
#去掉一些超过选择范围的地点
if v[0]>=len(img_desc_0):
continue
if v[0] in pos_dict.keys():
pos_dict[v[0]].append(v[1])
else:
pos_dict[v[0]] = [v[1]]
out_save = []
recall = np.array([0.]*25)
for v in tqdm(pos_dict.keys()):
print('v',v)
test_dataset = seq2pc_test_recall(seq=seq,
v= v,
img_desc_0= img_desc_0,
img_desc_5= img_desc_5,
img_desc_10= img_desc_10,
img_desc_15= img_desc_15,
img_desc_cb= img_desc_cb,
velo_desc_0= velo_desc_0,
velo_desc_1= velo_desc_1,
velo_desc_2= velo_desc_2,
velo_desc_3= velo_desc_3,
velo_desc_4= velo_desc_4,
velo_desc_5= velo_desc_5,
velo_desc_6= velo_desc_6,
velo_desc_7= velo_desc_7,
)
batch_size=1024
test_loader = DataLoader(
dataset=test_dataset, batch_size=batch_size, shuffle=False, num_workers=0)
with torch.no_grad():
out_list=[]
for i_batch, sample_batch in tqdm(enumerate(test_loader), total=len(test_loader), desc="Test", leave=False):
input=torch.cat((sample_batch["img_descs_0"].unsqueeze(1),sample_batch["img_descs_5"].unsqueeze(1),sample_batch["img_descs_10"].unsqueeze(1),sample_batch["img_descs_15"].unsqueeze(1)),1)
input=input.to(device) #b*4*12*90
seq_contour_matrix, mask=model_mae(input)
seq_contour_matrix=torch.clamp(seq_contour_matrix,min=0.0,max=1.0)
#保存生成的轮廓矩阵
pad = (135, 135) # 在 最后1 维度上左侧补充 135 个 0，右侧补充 135 个 0
seq_contour_matrix=torch.nn.functional.pad(seq_contour_matrix, pad, mode='constant', value=0) #64, 1, 12, 360
img_contour_matrix=torch.nn.functional.pad(sample_batch["img_descs_0"].unsqueeze(1), pad, mode='constant', value=0) #64, 1, 12, 360
out, diff,out_cat = net(seq_contour_matrix.squeeze(1)-img_contour_matrix.squeeze(1).to(device=device),
img_contour_matrix.squeeze(1).to(device=device),
sample_batch["desc2_0"].to(device=device),
sample_batch["desc2_1"].to(device=device),
sample_batch["desc2_2"].to(device=device),
sample_batch["desc2_3"].to(device=device),
sample_batch["desc2_4"].to(device=device),
sample_batch["desc2_5"].to(device=device),
sample_batch["desc2_6"].to(device=device),
sample_batch["desc2_7"].to(device=device),
)
out_list.append(out)
out_list=torch.cat(out_list,dim=0)
out_list = out_list.cpu().numpy()
ids = np.argsort(-out_list)
o = [v]
o += ids[:25].tolist()
for i in range(25):
if ids[i] in pos_dict[v]:
o+=[True]
else:
o+=[False]
for i in range(25):
if ids[i] in pos_dict[v]:
recall[i:] += 1
break
out_save.append(o)
if not os.path.exists('result'):
os.mkdir('result')
np.savetxt(os.path.join('result', seq+'_recall_retrieval.txt'), out_save, fmt='%d')
recall /= len(pos_dict.keys())
print(recall)
np.savetxt(os.path.join('result', seq+'_recall_scores.txt'), recall)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--seq', default='00',
help='Sequence to eval. ')
parser.add_argument('--dataset', default="kitti",
help="Dataset ")
parser.add_argument('--model', default="./checkpoints/00.ckpt",
help='Model file. ')
parser.add_argument('--model_mae', default="./checkpoints/00_mae.ckpt",
help='Model file. ')
parser.add_argument('--velo_desc_folder', default='./lidar_desc',
help='folder of lidar descriptors. ')
parser.add_argument('--img_desc_folder', default='./img_desc',
help='folder of image descriptors. ')
parser.add_argument('--pose_file', default="./pose_kitti/00.txt",
help='Pose file (eval_type=recall). ')
parser.add_argument('--eval_type', default="recall",
help='Type of evaluation (f1 or recall). [default: f1]')
cfg = parser.parse_args()
if cfg.dataset == "kitti" and cfg.eval_type == "recall":
recall(seq=cfg.seq, model_file=cfg.model, model_mae_file=cfg.model_mae,
velo_desc_folder=cfg.velo_desc_folder,
img_desc_folder=cfg.img_desc_folder,
pose_file=cfg.pose_file)
else:
print("Error")