-
Notifications
You must be signed in to change notification settings - Fork 6
/
train-unused.py
executable file
·469 lines (405 loc) · 19.8 KB
/
train-unused.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
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
from pathlib import Path
import argparse
import random
import numpy as np
import matplotlib.cm as cm
import torch
import torch.nn as nn
from torch.autograd import Variable
from load_data import SparseDataset
import os
import torch.multiprocessing
from tqdm import tqdm
# from models.matching import Matching
from models.utils import (compute_pose_error, compute_epipolar_error,
estimate_pose, make_matching_plot,
error_colormap, AverageTimer, pose_auc, read_image,
rotate_intrinsics, rotate_pose_inplane,
scale_intrinsics)
from models.superpoint import SuperPoint
from models.superglue import SuperGlue
from models.matchingForTraining import MatchingForTraining
torch.set_grad_enabled(True)
torch.multiprocessing.set_sharing_strategy('file_system')
# os.environ["CUDA_VISIBLE_DEVICES"]="0"
# torch.multiprocessing.set_start_method("spawn")
# torch.cuda.set_device(0)
# try:
# torch.multiprocessing.set_start_method('spawn')
# except RuntimeError:
# pass
parser = argparse.ArgumentParser(
description='Image pair matching and pose evaluation with SuperGlue',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument(
'--viz', action='store_true',
help='Visualize the matches and dump the plots')
parser.add_argument(
'--eval', action='store_true',
help='Perform the evaluation'
' (requires ground truth pose and intrinsics)')
parser.add_argument(
'--superglue', choices={'indoor', 'outdoor'}, default='indoor',
help='SuperGlue weights')
parser.add_argument(
'--max_keypoints', type=int, default=1024,
help='Maximum number of keypoints detected by Superpoint'
' (\'-1\' keeps all keypoints)')
parser.add_argument(
'--keypoint_threshold', type=float, default=0.005,
help='SuperPoint keypoint detector confidence threshold')
parser.add_argument(
'--nms_radius', type=int, default=4,
help='SuperPoint Non Maximum Suppression (NMS) radius'
' (Must be positive)')
parser.add_argument(
'--sinkhorn_iterations', type=int, default=20,
help='Number of Sinkhorn iterations performed by SuperGlue')
parser.add_argument(
'--match_threshold', type=float, default=0.2,
help='SuperGlue match threshold')
parser.add_argument(
'--resize', type=int, nargs='+', default=[640, 480],
help='Resize the input image before running inference. If two numbers, '
'resize to the exact dimensions, if one number, resize the max '
'dimension, if -1, do not resize')
parser.add_argument(
'--resize_float', action='store_true',
help='Resize the image after casting uint8 to float')
parser.add_argument(
'--cache', action='store_true',
help='Skip the pair if output .npz files are already found')
parser.add_argument(
'--show_keypoints', action='store_true',
help='Plot the keypoints in addition to the matches')
parser.add_argument(
'--fast_viz', action='store_true',
help='Use faster image visualization based on OpenCV instead of Matplotlib')
parser.add_argument(
'--viz_extension', type=str, default='png', choices=['png', 'pdf'],
help='Visualization file extension. Use pdf for highest-quality.')
parser.add_argument(
'--opencv_display', action='store_true',
help='Visualize via OpenCV before saving output images')
parser.add_argument(
'--eval_pairs_list', type=str, default='assets/scannet_sample_pairs_with_gt.txt',
help='Path to the list of image pairs for evaluation')
parser.add_argument(
'--shuffle', action='store_true',
help='Shuffle ordering of pairs before processing')
parser.add_argument(
'--max_length', type=int, default=-1,
help='Maximum number of pairs to evaluate')
parser.add_argument(
'--eval_input_dir', type=str, default='assets/scannet_sample_images/',
help='Path to the directory that contains the images')
parser.add_argument(
'--eval_output_dir', type=str, default='dump_match_pairs/',
help='Path to the directory in which the .npz results and optional,'
'visualizations are written')
parser.add_argument(
'--learning_rate', type=int, default=0.001,
help='Learning rate')
parser.add_argument(
'--batch_size', type=int, default=1,
help='batch_size')
parser.add_argument(
'--train_path', type=str, default='/dev/shm/MSCOCO_50/', # MSCOCO2014_yingxin
help='Path to the directory of training imgs.')
parser.add_argument(
'--nfeatures', type=int, default=80,
help='Number of feature points to be extracted initially, in each img.')
parser.add_argument(
'--epoch', type=int, default=20,
help='Number of epoches')
if __name__ == '__main__':
opt = parser.parse_args()
print(opt)
assert not (opt.opencv_display and not opt.viz), 'Must use --viz with --opencv_display'
assert not (opt.opencv_display and not opt.fast_viz), 'Cannot use --opencv_display without --fast_viz'
assert not (opt.fast_viz and not opt.viz), 'Must use --viz with --fast_viz'
assert not (opt.fast_viz and opt.viz_extension == 'pdf'), 'Cannot use pdf extension with --fast_viz'
with open(opt.eval_pairs_list, 'r') as f: # eval
pairs = [l.split() for l in f.readlines()]
if opt.max_length > -1:
pairs = pairs[0:np.min([len(pairs), opt.max_length])]
if opt.shuffle:
random.Random(0).shuffle(pairs)
if not all([len(p) == 38 for p in pairs]):
raise ValueError(
'All pairs should heval_ave ground truth info for evaluation.'
'File \"{}\" needs 38 valid entries per row'.format(opt.eval_pairs_list))
# Create the output directories if they do not exist already.
eval_input_dir = Path(opt.eval_input_dir)
print('Looking for data in directory \"{}\"'.format(eval_input_dir))
eval_output_dir = Path(opt.eval_output_dir)
eval_output_dir.mkdir(exist_ok=True, parents=True)
print('Will write matches to directory \"{}\"'.format(eval_output_dir))
print('Will write evaluation results',
'to directory \"{}\"'.format(eval_output_dir))
if opt.viz:
print('Will write visualization images to',
'directory \"{}\"'.format(eval_output_dir))
timer = AverageTimer(newline=True)
config = {
'superpoint': {
'nms_radius': opt.nms_radius,
'keypoint_threshold': opt.keypoint_threshold,
'max_keypoints': opt.max_keypoints
},
'superglue': {
'weights': opt.superglue,
'sinkhorn_iterations': opt.sinkhorn_iterations,
'match_threshold': opt.match_threshold,
}
}
train_set = SparseDataset(opt.train_path, opt.nfeatures)
train_loader = torch.utils.data.DataLoader(dataset=train_set, shuffle=False, batch_size=opt.batch_size, drop_last=True)
superpoint = SuperPoint(config.get('superpoint', {}))
superglue = SuperGlue(config.get('superglue', {}))
if torch.cuda.is_available():
superpoint.cuda()
superglue.cuda()
else:
print("### CUDA not available ###")
optimizer = torch.optim.Adam(superglue.parameters(), lr=opt.learning_rate)
mean_loss = []
for epoch in range(1, opt.epoch+1):
epoch_loss = 0
superglue.double().train()
# train_loader = tqdm(train_loader)
for i, pred in enumerate(train_loader):
for k in pred:
if k != 'file_name':
if type(pred[k]) == torch.Tensor:
pred[k] = Variable(pred[k].cuda())
else:
pred[k] = Variable(torch.stack(pred[k]).cuda())
data = superglue(pred)
for k, v in pred.items():
pred[k] = v[0]
pred = {**pred, **data}
if pred['skip_train'] == True: # image has no keypoint
continue
superglue.zero_grad()
Loss = pred['loss']
epoch_loss += Loss
mean_loss.append(Loss) # every 10 pairs
Loss.backward()
optimizer.step()
if (i+1) % 10 == 0:
print ('Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}'
.format(epoch, opt.epoch, i+1, len(train_loader), torch.mean(torch.stack(mean_loss)).item())) # Loss.item()
mean_loss = []
model_out_path = "model_epoch_{}.pth".format(epoch)
torch.save(superglue, model_out_path)
print("Checkpoint saved to {}".format(model_out_path))
# timer.print('Finished pair {:5} of {:5}'.format(i, len(pairs)))
### eval ###
superglue.eval()
with torch.no_grad():
# val_bar = tqdm(val_loader)
for i, pair in enumerate(pairs):
name0, name1 = pair[:2]
stem0, stem1 = Path(name0).stem, Path(name1).stem
matches_path = eval_output_dir / '{}_{}_matches.npz'.format(stem0, stem1)
eval_path = eval_output_dir / '{}_{}_evaluation.npz'.format(stem0, stem1)
viz_path = eval_output_dir / '{}_{}_matches.{}'.format(stem0, stem1, opt.viz_extension)
viz_eval_path = eval_output_dir / \
'{}_{}_evaluation.{}'.format(stem0, stem1, opt.viz_extension)
# Handle --cache logic.
do_match = True
do_viz = opt.viz
do_viz_eval = opt.viz
if opt.cache:
if matches_path.exists():
try:
results = np.load(matches_path)
except:
raise IOError('Cannot load matches .npz file: %s' %
matches_path)
kpts0, kpts1 = results['keypoints0'], results['keypoints1']
matches, conf = results['matches'], results['match_confidence']
do_match = False
if eval_path.exists():
try:
results = np.load(eval_path)
except:
raise IOError('Cannot load eval .npz file: %s' % eval_path)
err_R, err_t = results['error_R'], results['error_t']
precision = results['precision']
matching_score = results['matching_score']
num_correct = results['num_correct']
epi_errs = results['epipolar_errors']
do_eval = False
if opt.viz and viz_path.exists():
do_viz = False
if opt.viz and viz_eval_path.exists():
do_viz_eval = False
timer.update('load_cache')
if not (do_match or do_eval or do_viz or do_viz_eval):
timer.print('Finished pair {:5} of {:5}'.format(i, len(pairs)))
continue
# If a rotation integer is provided (e.g. from EXIF data), use it:
if len(pair) >= 5:
rot0, rot1 = int(pair[2]), int(pair[3])
else:
rot0, rot1 = 0, 0
# Load the image pair.
image0, inp0, scales0 = read_image(
eval_input_dir / name0, opt.resize, rot0, opt.resize_float)
image1, inp1, scales1 = read_image(
eval_input_dir / name1, opt.resize, rot1, opt.resize_float)
if image0 is None or image1 is None:
print('Problem reading image pair: {} {}'.format(
eval_input_dir/name0, eval_input_dir/name1))
exit(1)
timer.update('load_image')
if do_match:
# Perform the matching.
# pred_eval = matching({'image0': inp0, 'image1': inp1})
data = {'image0': inp0, 'image1': inp1}
pred0 = superpoint({'image': data['image0']})
pred = {**pred, **{k+'0': v for k, v in pred0.items()}}
pred1 = superpoint({'image': data['image1']})
pred = {**pred, **{k+'1': v for k, v in pred1.items()}}
data = {**data, **pred}
for k in data:
if isinstance(data[k], (list, tuple)):
data[k] = torch.stack(data[k])
pred = {**pred, **superglue(data)}
pred.pop('skip_train', None)
pred.pop('loss', None)
pred_eval = {k: v[0].cpu().numpy() for k, v in pred_eval.items()}
kpts0, kpts1 = pred_eval['keypoints0'], pred_eval['keypoints1']
matches, conf = pred_eval['matches0'], pred_eval['matching_scores0']
timer.update('matcher_eval')
# Write the matches to disk.
out_matches = {'keypoints0': kpts0, 'keypoints1': kpts1,
'matches': matches, 'match_confidence': conf}
np.savez(str(matches_path), **out_matches)
# Keep the matching keypoints.
valid = matches > -1
mkpts0 = kpts0[valid]
mkpts1 = kpts1[matches[valid]]
mconf = conf[valid]
# Estimate the pose and compute the pose error.
assert len(pair) == 38, 'Pair does not have ground truth info'
K0 = np.array(pair[4:13]).astype(float).reshape(3, 3)
K1 = np.array(pair[13:22]).astype(float).reshape(3, 3)
T_0to1 = np.array(pair[22:]).astype(float).reshape(4, 4)
# Scale the intrinsics to resized image.
K0 = scale_intrinsics(K0, scales0)
K1 = scale_intrinsics(K1, scales1)
# Update the intrinsics + extrinsics if EXIF rotation was found.
if rot0 != 0 or rot1 != 0:
cam0_T_w = np.eye(4)
cam1_T_w = T_0to1
if rot0 != 0:
K0 = rotate_intrinsics(K0, image0.shape, rot0)
cam0_T_w = rotate_pose_inplane(cam0_T_w, rot0)
if rot1 != 0:
K1 = rotate_intrinsics(K1, image1.shape, rot1)
cam1_T_w = rotate_pose_inplane(cam1_T_w, rot1)
cam1_T_cam0 = cam1_T_w @ np.linalg.inv(cam0_T_w)
T_0to1 = cam1_T_cam0
epi_errs = compute_epipolar_error(mkpts0, mkpts1, T_0to1, K0, K1)
correct = epi_errs < 5e-4
num_correct = np.sum(correct)
precision = np.mean(correct) if len(correct) > 0 else 0
matching_score = num_correct / len(kpts0) if len(kpts0) > 0 else 0
thresh = 1. # In pixels relative to resized image size.
ret = estimate_pose(mkpts0, mkpts1, K0, K1, thresh)
if ret is None:
err_t, err_R = np.inf, np.inf
else:
R, t, inliers = ret
err_t, err_R = compute_pose_error(T_0to1, R, t)
# Write the evaluation results to disk.
out_eval = {'error_t': err_t,
'error_R': err_R,
'precision': precision,
'matching_score': matching_score,
'num_correct': num_correct,
'epipolar_errors': epi_errs}
np.savez(str(eval_path), **out_eval)
timer.update('eval')
if do_viz:
# Visualize the matches.
color = cm.jet(mconf)
text = [
'SuperGlue',
'Keypoints: {}:{}'.format(len(kpts0), len(kpts1)),
'Matches: {}'.format(len(mkpts0)),
]
if rot0 != 0 or rot1 != 0:
text.append('Rotation: {}:{}'.format(rot0, rot1))
# Display extra parameter info.
k_thresh = matching.superpoint.config['keypoint_threshold']
m_thresh = matching.superglue.config['match_threshold']
small_text = [
'Keypoint Threshold: {:.4f}'.format(k_thresh),
'Match Threshold: {:.2f}'.format(m_thresh),
'Image Pair: {}:{}'.format(stem0, stem1),
]
make_matching_plot(
image0, image1, kpts0, kpts1, mkpts0, mkpts1, color,
text, viz_path, opt.show_keypoints,
opt.fast_viz, opt.opencv_display, 'Matches', small_text)
timer.update('viz_match')
if do_viz_eval:
# Visualize the evaluation results for the image pair.
color = np.clip((epi_errs - 0) / (1e-3 - 0), 0, 1)
color = error_colormap(1 - color)
deg, delta = ' deg', 'Delta '
if not opt.fast_viz:
deg, delta = '°', '$\\Delta$'
e_t = 'FAIL' if np.isinf(err_t) else '{:.1f}{}'.format(err_t, deg)
e_R = 'FAIL' if np.isinf(err_R) else '{:.1f}{}'.format(err_R, deg)
text = [
'SuperGlue',
'{}R: {}'.format(delta, e_R), '{}t: {}'.format(delta, e_t),
'inliers: {}/{}'.format(num_correct, (matches > -1).sum()),
]
if rot0 != 0 or rot1 != 0:
text.append('Rotation: {}:{}'.format(rot0, rot1))
# Display extra parameter info (only works with --fast_viz).
k_thresh = matching.superpoint.config['keypoint_threshold']
m_thresh = matching.superglue.config['match_threshold']
small_text = [
'Keypoint Threshold: {:.4f}'.format(k_thresh),
'Match Threshold: {:.2f}'.format(m_thresh),
'Image Pair: {}:{}'.format(stem0, stem1),
]
make_matching_plot(
image0, image1, kpts0, kpts1, mkpts0,
mkpts1, color, text, viz_eval_path,
opt.show_keypoints, opt.fast_viz,
opt.opencv_display, 'Relative Pose', small_text)
timer.update('viz_eval')
timer.print('Finished pair {:5} of {:5}'.format(i, len(pairs)))
# Collate the results into a final table and print to terminal.
pose_errors = []
precisions = []
matching_scores = []
for pair in pairs:
name0, name1 = pair[:2]
stem0, stem1 = Path(name0).stem, Path(name1).stem
eval_path = eval_output_dir / \
'{}_{}_evaluation.npz'.format(stem0, stem1)
results = np.load(eval_path)
pose_error = np.maximum(results['error_t'], results['error_R'])
pose_errors.append(pose_error)
precisions.append(results['precision'])
matching_scores.append(results['matching_score'])
thresholds = [5, 10, 20]
aucs = pose_auc(pose_errors, thresholds)
aucs = [100.*yy for yy in aucs]
prec = 100.*np.mean(precisions)
ms = 100.*np.mean(matching_scores)
print ('Epoch [{}/{}]'.format(epoch, opt.epoch))
print("===> Epoch {} Complete: Avg. Loss: {:.4f}".format(epoch, epoch_loss ))
print('Evaluation Results (mean over {} pairs):'.format(len(pairs)))
print('AUC@5\t AUC@10\t AUC@20\t Prec\t MScore\t')
print('{:.2f}\t {:.2f}\t {:.2f}\t {:.2f}\t {:.2f}\t'.format(
aucs[0], aucs[1], aucs[2], prec, ms))