def __init__(self):
self.matching = Matching(self.config).eval().to(self.device)
self.keys = ["keypoints", "scores", "descriptors"]
### Anchor Frame and associated data
self.anchor_frame_tensor = None
self.anchor_data = None
self.anchor_frame = None
self.anchor_image_id = None
device = 'cuda' if torch.cuda.is_available(
) and not opt.force_cpu else 'cpu'
print('Running inference on device \"{}\"'.format(device))
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,
}
}
matching = Matching(config).eval().to(device)
keys = ['keypoints', 'scores', 'descriptors']
vs = VideoStreamer(opt.input, opt.resize, opt.skip, opt.image_glob,
opt.max_length)
frame, ret = vs.next_frame()
assert ret, 'Error when reading the first frame (try different --input?)'
frame_tensor = frame2tensor(frame, device)
last_data = matching.superpoint({'image': frame_tensor})
last_data = {k + '0': last_data[k] for k in keys}
last_data['image0'] = frame_tensor
last_frame = frame
last_image_id = 0
if opt.output_dir is not None:
def get_similarity(args, name, device='cpu'):
args.set_match(name)
matching = Matching(args.config).eval().to(device) # matching model
return matching(args.data)
# self.anchor_features.append(sc)
# self.anchor_features.append(desc)
def get_similarity(args, name, device='cpu'):
args.set_match(name)
matching = Matching(args.config).eval().to(device) # matching model
return matching(args.data)
if __name__ == '__main__':
args = Arg(scene_mode='outdoor')
args.set_anchor('zju1.jpg')
args.set_match('zju2.jpg')
# print(get_similarity(args, 'zju2.jpg'))
m = Matching(args.config).eval()
# print(m(args.data))
# ms=torch.jit.trace(m,args.data)
# pt2onnx(m,args.data)
# torch.set_grad_enabled(False)
# device = 'cuda' if torch.cuda.is_available() and not opt.force_cpu else 'cpu'
# name0 = 'anchor.jpg'
# args.set_anchor(name0)
# name1 = 'to_match.jpg'
# name2 = 'to_match2.jpg'
# name3 = 'to_match3.jpg'
# name4 = 'to_match4.jpg'
# print(get_similarity(args, name1))
# print(get_similarity(args, name2))
# # print(get_similarity(args, name3))
def on_train_epoch_end(self, outputs) -> None:
for eval_data in self.hparams.eval:
with open(f'{ROOT_PATH}/' + eval_data.pairs_list, 'r') as f:
pairs = [l.split() for l in f.readlines()]
if eval_data.max_length > -1:
pairs = pairs[0:np.min([len(pairs), eval_data.max_length])]
if eval_data.shuffle:
random.Random(0).shuffle(pairs)
if not all([len(p) == 38 for p in pairs]):
raise ValueError(
'All pairs should have ground truth info for evaluation.'
'File \"{}\" needs 38 valid entries per row'.format(
eval_data.pairs_list))
# Load the SuperPoint and SuperGlue models.
device = 'cuda' if torch.cuda.is_available() else 'cpu'
print('Running inference on device \"{}\"'.format(device))
config = {
'superpoint': {
'nms_radius': eval_data.nms_radius,
'keypoint_threshold': eval_data.keypoint_threshold,
'max_keypoints': eval_data.max_keypoints
},
'superglue': self.hparams.model.superglue,
}
matching = Matching(config).eval().to(device)
matching.superglue.load_state_dict(self.superglue.state_dict())
# Create the output directories if they do not exist already.
data_dir = Path(f'{ROOT_PATH}/' + eval_data.data_dir)
# moving_dir = Path(f'{ROOT_PATH}/' + 'data/ScanNet/test_subset')
print('Looking for data in directory \"{}\"'.format(data_dir))
results_dir = Path(os.getcwd() + '/' + eval_data.results_dir)
results_dir.mkdir(exist_ok=True, parents=True)
print('Will write matches to directory \"{}\"'.format(results_dir))
timer = AverageTimer(newline=True)
for i, pair in enumerate(pairs):
name0, name1 = pair[:2]
stem0, stem1 = Path(name0).stem, Path(name1).stem
matches_path = results_dir / '{}_{}_matches.npz'.format(
stem0, stem1)
eval_path = results_dir / '{}_{}_evaluation.npz'.format(
stem0, stem1)
viz_path = results_dir / '{}_{}_matches.{}'.format(
stem0, stem1, self.hparams.exp.viz_extension)
viz_eval_path = results_dir / \
'{}_{}_evaluation.{}'.format(stem0, stem1, self.hparams.exp.viz_extension)
# Handle --cache logic.
do_match = True
do_eval = True
do_viz = self.hparams.exp.viz
do_viz_eval = self.hparams.exp.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 opt.eval and 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 opt.eval 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(data_dir / name0,
eval_data.resize, rot0,
eval_data.resize_float)
image1, inp1, scales1 = read_image(data_dir / name1,
eval_data.resize, rot1,
eval_data.resize_float)
# Moving
# os.makedirs(os.path.dirname(moving_dir / name0), exist_ok=True)
# os.makedirs(os.path.dirname(moving_dir / name1), exist_ok=True)
# shutil.copy(data_dir / name0, moving_dir / name0)
# shutil.copy(data_dir / name1, moving_dir / name1)
if image0 is None or image1 is None:
print('Problem reading image pair: {} {}'.format(
data_dir / name0, data_dir / name1))
exit(1)
timer.update('load_image')
if do_match:
# Perform the matching.
with torch.no_grad():
pred = matching({
'image0': inp0.cuda(),
'image1': inp1.cuda()
})
pred_np = {}
for (k, v) in pred.items():
if isinstance(v, list):
pred_np[k] = v[0].cpu().numpy()
elif isinstance(v, torch.Tensor):
pred_np[k] = v[0].cpu().numpy()
pred = pred_np
# pred = {k: v[0].cpu().numpy() for k, v in pred.items() if isinstance(v, torch.Tensor)}
kpts0, kpts1 = pred['keypoints0'], pred['keypoints1']
matches, conf = pred['matches0'], pred['matching_scores0']
timer.update('matcher')
# 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]
if do_eval:
# 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))
#
# make_matching_plot(
# image0, image1, kpts0, kpts1, mkpts0, mkpts1, color,
# text, viz_path, stem0, stem1, opt.show_keypoints,
# opt.fast_viz, opt.opencv_display, 'Matches')
#
# 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))
#
# make_matching_plot(
# image0, image1, kpts0, kpts1, mkpts0,
# mkpts1, color, text, viz_eval_path,
# stem0, stem1, opt.show_keypoints,
# opt.fast_viz, opt.opencv_display, 'Relative Pose')
#
# 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 = results_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('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))
self.log(f'{eval_data.name}/AUC_5',
aucs[0],
on_epoch=True,
on_step=False)
self.log(f'{eval_data.name}/AUC_10',
aucs[1],
on_epoch=True,
on_step=False)
self.log(f'{eval_data.name}/AUC_20',
aucs[2],
on_epoch=True,
on_step=False)
self.log(f'{eval_data.name}/Prec',
prec,
on_epoch=True,
on_step=False)
self.log(f'{eval_data.name}/MScore',
ms,
on_epoch=True,
on_step=False)
class SuperMatcher:
resize = [640, 480]
superglue = "outdoor"
max_keypoints = -1
nms_radius = 4
keypoint_threshold = 0.1
sinkhorn_iterations = 20
match_threshold = 0.4
device = "cuda" if torch.cuda.is_available() else "cpu"
show_keypoints = False
config = {
"superpoint": {
"nms_radius": nms_radius,
"keypoint_threshold": keypoint_threshold,
"max_keypoints": max_keypoints,
},
"superglue": {
"weights": superglue,
"sinkhorn_iterations": sinkhorn_iterations,
"match_threshold": match_threshold,
},
}
def __init__(self):
self.matching = Matching(self.config).eval().to(self.device)
self.keys = ["keypoints", "scores", "descriptors"]
### Anchor Frame and associated data
self.anchor_frame_tensor = None
self.anchor_data = None
self.anchor_frame = None
self.anchor_image_id = None
def set_anchor(self, frame):
# Frame will be divided by 255
self.anchor_frame_tensor = frame2tensor(frame, self.device)
self.anchor_data = self.matching.superpoint(
{"image": self.anchor_frame_tensor})
self.anchor_data = {k + "0": self.anchor_data[k] for k in self.keys}
self.anchor_data["image0"] = self.anchor_frame_tensor
self.anchor_frame = frame
self.anchor_image_id = 0
def process(self, frame):
if self.anchor_frame_tensor is None:
print("Please set anchor frame first...")
return None
frame_tensor = frame2tensor(frame, self.device)
pred = self.matching({**self.anchor_data, "image1": frame_tensor})
kpts0 = self.anchor_data["keypoints0"][0].cpu().numpy()
kpts1 = pred["keypoints1"][0].cpu().numpy()
matches = pred["matches0"][0].cpu().numpy()
confidence = pred["matching_scores0"][0].cpu().numpy()
valid = matches > -1
mkpts0 = kpts0[valid]
mkpts1 = kpts1[matches[valid]]
color = cm.jet(confidence[valid])
text = [
"SuperGlue",
"Keypoints: {}:{}".format(len(kpts0), len(kpts1)),
"Matches: {}".format(len(mkpts0)),
]
k_thresh = self.matching.superpoint.config["keypoint_threshold"]
m_thresh = self.matching.superglue.config["match_threshold"]
small_text = [
"Keypoint Threshold: {:.4f}".format(k_thresh),
"Match Threshold: {:.2f}".format(m_thresh),
]
out = make_matching_plot_fast(
self.anchor_frame,
frame,
kpts0,
kpts1,
mkpts0,
mkpts1,
color,
text,
path=None,
show_keypoints=self.show_keypoints,
small_text=small_text,
)
return out, mkpts0, mkpts1
# Load the SuperPoint and SuperGlue models.
device = 'cuda'
print('Running inference on device \"{}\"'.format(device))
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,
}
}
matching = Matching(config).eval().to(device)
superpoint = SuperPoint(config.get('superpoint', {})).eval().to(device)
# Create the output directories if they do not exist already.
input_dir = Path(opt.input_dir)
print('Looking for data in directory \"{}\"'.format(input_dir))
output_dir = Path(opt.output_dir)
output_dir.mkdir(exist_ok=True, parents=True)
print('Will write matches to directory \"{}\"'.format(output_dir))
if opt.viz:
print('Will write visualization images to',
'directory \"{}\"'.format(output_dir))
image_tensor = {}
keypoints_tensor = {}
response = {}
matches_mvg = {}
if opt.ensemble:
model1 = gdes.APGeM(config["APGeM"]["Path1"])
model2 = gdes.APGeM(config["APGeM"]["Path2"])
print("Model load Done")
if opt.ensemble:
knn = gdes.make_cache(DB_cache_name, pickle_name, model,
imgDataLoader, model1, model2)
else:
knn = gdes.make_cache(DB_cache_name, pickle_name, model,
imgDataLoader)
print("===>Predicting")
matching = Matching(
config["SuperGlue"]["config"]).eval().to(device)
for i in tqdm(range(len(file_list))):
max_match = 0
max_index = 0
max_query = 0
sub_file_list = sorted(glob.glob(file_list[i] + "*_L.png"))
for tt in range(len(sub_file_list)):
print(sub_file_list[tt])
odmetry = np.loadtxt(os.path.join(odmetry_list[i]),
delimiter=' ').reshape(-1, 4, 4)
query_path = sub_file_list[tt]
device = 'cuda' if torch.cuda.is_available(
) and not opt.force_cpu else 'cpu'
print('Running inference on device \"{}\"'.format(device))
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,
}
}
matching = Matching(config, opt).eval().to(device)
keys = ['keypoints', 'scores', 'descriptors']
vs = VideoStreamer(opt.input, opt.resize, opt.skip, opt.image_glob,
opt.max_length)
if opt.regionnetvlad:
vs.rgb = True
frame, ret = vs.next_frame()
frame = cv2.resize(frame, (640, 480))
# print('frameshape', frame.shape)
assert ret, 'Error when reading the first frame (try different --input?)'
frame_tensor = frame2tensor(frame, device)
if opt.regionnetvlad:
last_data = matching.image_transformer(frame)
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,
'descriptor_dim': detector_dims[opt.detector],
'detector': opt.detector
}
}
matching = Matching(config, is_train=False).eval().to(device)
# Create the output directories if they do not exist already.
data_dir = Path(opt.data_dir)
print('Looking for data in directory \"{}\"'.format(data_dir))
results_dir = Path(opt.results_dir)
results_dir.mkdir(exist_ok=True, parents=True)
print('Will write matches to directory \"{}\"'.format(results_dir))
if opt.eval:
print('Will write evaluation results',
'to directory \"{}\"'.format(results_dir))
if opt.viz:
print('Will write visualization images to',
'directory \"{}\"'.format(results_dir))
timer = AverageTimer(newline=True)
