-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathpredict.py
209 lines (166 loc) · 9.27 KB
/
predict.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
import argparse
import torch
import os
import datetime
from tensorboardX import SummaryWriter
import sys
from dataset import find_dataset_def
import torch.backends.cudnn as cudnn
from networks.casmvs import CascadeMVSNet
from networks.ucs import UCSNet
# from networks.casred import Infer_CascadeREDNet
from networks.stsat import ST_SatMVS, Infer_CascadeREDNet
from torch.utils.data import DataLoader
import torch.nn as nn
import time
from tools.utils import *
from dataset.data_io import save_pfm
import matplotlib.pyplot as plt
cudnn.benchmark = True
parser = argparse.ArgumentParser(description='A PyTorch Implementation')
parser.add_argument('--model', default="samsat", help='select model', choices=['SAMsat', 'red', "casmvs", "ucs"])
parser.add_argument('--geo_model', default="rpc", help='select dataset', choices=["rpc", "pinhole"])
parser.add_argument('--use_qc', default=False, help="whether to use Quaternary Cubic Form for RPC warping.")
parser.add_argument('--dataset_root', default='/remote-home/Cs_ai_qj_new/chenziyang/MVS/MVSrs/open_dataset_rpc/test', help='dataset root')
parser.add_argument('--loadckpt', default="./checkpoints/samsat/rpc/model_000012.ckpt",
help='load a specific checkpoint')
# input parameters
parser.add_argument('--view_num', type=int, default=3, help='Number of images.')
parser.add_argument('--ref_view', type=int, default=2)
parser.add_argument('--batch_size', type=int, default=1, help='train batch size')
# Cascade parameters
parser.add_argument('--ndepths', type=str, default="64,32,8", help='ndepths')
parser.add_argument('--min_interval', type=float, default=2.5, help='min_interval in the bottom stage')
parser.add_argument('--depth_inter_r', type=str, default="4,2,1", help='depth_intervals_ratio')
parser.add_argument('--lamb', type=float, default=1.5, help="lamb in ucs-net")
parser.add_argument('--cr_base_chs', type=str, default="8,8,8", help='cost regularization base channels')
parser.add_argument('--gpu_id', type=str, default="2")
# parse arguments and check
args = parser.parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_id
# print(args.geo_model)
# print(args.dataset_root)
assert args.geo_model in args.dataset_root, Exception("set the wrong data root")
# assert args.geo_model in args.loadckpt, Exception("set the wrong checkpoint")
# assert args.model in args.loadckpt, Exception("set the wrong checkpoint")
def predict():
print("argv:", sys.argv[1:])
print_args(args)
# dataset, dataloader
MVSDataset = find_dataset_def(args.geo_model)
# pre_dataset = MVSDataset(args.dataset_root, "pred", args.view_num, ref_view=args.ref_view, use_qc=args.use_qc)
pre_dataset = MVSDataset(args.dataset_root, "test", args.view_num, ref_view=args.ref_view, use_qc=args.use_qc)
Pre_ImgLoader = DataLoader(pre_dataset, args.batch_size, shuffle=False, num_workers=0, drop_last=False)
if args.model == "casmvs":
model = CascadeMVSNet(min_interval=args.min_interval,
ndepths=[int(nd) for nd in args.ndepths.split(",") if nd],
depth_interals_ratio=[float(d_i) for d_i in args.depth_inter_r.split(",") if d_i],
cr_base_chs=[int(ch) for ch in args.cr_base_chs.split(",") if ch],
geo_model=args.geo_model, use_qc=args.use_qc)
print("===============> Model: Cascade MVS Net ===========>")
elif args.model == "ucs":
model = UCSNet(lamb=args.lamb, stage_configs=[int(nd) for nd in args.ndepths.split(",") if nd],
base_chs=[int(ch) for ch in args.cr_base_chs.split(",") if ch],
geo_model=args.geo_model, use_qc=args.use_qc)
print("===============> Model: UCS-Net ===========>")
elif args.model == "red":
model = Infer_CascadeREDNet(min_interval=args.min_interval,
ndepths=[int(nd) for nd in args.ndepths.split(",") if nd],
depth_interals_ratio=[float(d_i) for d_i in args.depth_inter_r.split(",") if d_i],
cr_base_chs=[int(ch) for ch in args.cr_base_chs.split(",") if ch],
geo_model=args.geo_model, use_qc=args.use_qc)
print("===============> Model: Cascade RED Net ===========>")
elif args.model == "samsat":
model = ST_SatMVS(min_interval=args.min_interval,
ndepths=[int(nd) for nd in args.ndepths.split(",") if nd],
depth_interals_ratio=[float(d_i) for d_i in args.depth_inter_r.split(",") if d_i],
cr_base_chs=[int(ch) for ch in args.cr_base_chs.split(",") if ch],
geo_model=args.geo_model, use_qc=args.use_qc)
print("===============> Model: Our network ===========>")
else:
raise Exception("{}? Not implemented yet!".format(args.model))
model = nn.DataParallel(model)
model.cuda()
# load checkpoint file specified by args.loadckpt
print("loading model {}".format(args.loadckpt))
state_dict = torch.load(args.loadckpt)
model.load_state_dict(state_dict['model'])
print('Number of model parameters: {}'.format(sum([p.data.nelement() for p in model.parameters()])))
# create output folder
# output_folder = os.path.join(args.dataset_root, 'mvs_results')
output_folder = os.path.join('./mvs_results')
if not os.path.isdir(output_folder):
os.mkdir(output_folder)
avg_test_scalars = DictAverageMeter()
t0 = time.time()
idx = 0
total_time = 0
for batch_idx, sample in enumerate(Pre_ImgLoader):
bview = sample['out_view'][0]
bname = sample['out_name'][0]
start_time = time.time()
scalar_outputs, image_outputs = predict_sample(model, sample)
avg_test_scalars.update(scalar_outputs)
scalar_outputs = {k: float("{0:.6f}".format(v)) for k, v in scalar_outputs.items()}
total_time += time.time() - start_time
print("Iter {}/{}, {}, time = {:3f}, test results = {}".format(batch_idx, len(Pre_ImgLoader), bname, time.time() - start_time, scalar_outputs))
# save results
depth_est = np.float32(np.squeeze(tensor2numpy(image_outputs["depth_est"])))
prob = np.float32(np.squeeze(tensor2numpy(image_outputs["photometric_confidence"])))
# TODO
depth_gt = sample['depth']['stage3']
mask = sample['mask']['stage3']
depth_gt = np.float32(np.squeeze(tensor2numpy(depth_gt)))
mask = (np.squeeze(tensor2numpy(mask))).astype(np.int)
depth_gt[mask < 0.5] = -999.0
# paths
output_folder2 = output_folder + ('/%s/' % bview)
if not os.path.exists(output_folder2):
os.mkdir(output_folder2)
if not os.path.exists(output_folder2 + '/prob/'):
os.mkdir(output_folder2 + '/prob/')
if not os.path.exists(output_folder2 + '/init/'):
os.mkdir(output_folder2 + '/init/')
if not os.path.exists(output_folder2 + '/prob/color/'):
os.mkdir(output_folder2 + '/prob/color/')
if not os.path.exists(output_folder2 + '/init/color/'):
os.mkdir(output_folder2 + '/init/color/')
init_depth_map_path = output_folder2 + ('/init/{}.pfm'.format(bname))
prob_map_path = output_folder2 + ('/prob/{}.pfm'.format(bname))
# save output
save_pfm(init_depth_map_path, depth_est)
save_pfm(prob_map_path, prob)
if args.geo_model == "pinhole":
depth_est = np.max(depth_est) - depth_est
# plt.imshow(depth_est)
# plt.show()
plt.imsave(output_folder2 + ('/init/color/{}.png'.format(bname)), depth_est, format='png')
plt.imsave(output_folder2 + ('/prob/color/{}_prob.png'.format(bname)), prob, format='png')
del scalar_outputs, image_outputs
# print("final, time = {:3f}, test results = {}".format(time.time() - t0, avg_test_scalars.mean()))
print("final, time = {:3f}, test results = {}".format(total_time, avg_test_scalars.mean()))
@make_nograd_func
def predict_sample(model, sample):
model.eval()
sample_cuda = tocuda(sample)
depth_gt_ms = sample_cuda["depth"]
mask_ms = sample_cuda["mask"]
num_stage = len([int(nd) for nd in args.ndepths.split(",") if nd])
depth_gt = depth_gt_ms["stage{}".format(num_stage)]
mask = mask_ms["stage{}".format(num_stage)]
outputs = model(sample_cuda["imgs"], sample_cuda["cam_para"], sample_cuda["depth_values"])
depth_est = outputs["stage3"]["depth"]
# depth_est = outputs["stage3"]["depth_filtered"]
photometric_confidence = outputs["stage3"]["photometric_confidence"]
image_outputs = {"depth_est": depth_est,
"photometric_confidence": photometric_confidence,
"ref_img": sample["imgs"][:, 0]}
scalar_outputs = {}
scalar_outputs["MAE"] = AbsDepthError_metrics(depth_est, depth_gt, mask > 0.5, 250.0)
scalar_outputs["RMSE"] = RMSE_metrics(depth_est, depth_gt, mask > 0.5, 250.0)
scalar_outputs["thres1.0m_error"] = Thres_metrics(depth_est, depth_gt, mask > 0.5, 1.0)
scalar_outputs["thres2.5m_error"] = Thres_metrics(depth_est, depth_gt, mask > 0.5, 2.5)
scalar_outputs["thres7.5m_error"] = Thres_metrics(depth_est, depth_gt, mask > 0.5, 7.5)
return tensor2float(scalar_outputs), image_outputs
if __name__ == '__main__':
predict()