def calibrate_neighbors(dataset,
config,
collate_fn,
keep_ratio=0.8,
samples_threshold=2000):
timer = Timer()
last_display = timer.total_time
# From config parameter, compute higher bound of neighbors number in a neighborhood
hist_n = int(np.ceil(4 / 3 * np.pi * (config.deform_radius + 1)**3))
neighb_hists = np.zeros((config.num_layers, hist_n), dtype=np.int32)
# Get histogram of neighborhood sizes i in 1 epoch max.
for i in range(len(dataset)):
timer.tic()
batched_input = collate_fn([dataset[i]],
config,
neighborhood_limits=[hist_n] * 5)
# update histogram
counts = [
torch.sum(neighb_mat < neighb_mat.shape[0], dim=1).numpy()
for neighb_mat in batched_input['neighbors']
]
hists = [np.bincount(c, minlength=hist_n)[:hist_n] for c in counts]
neighb_hists += np.vstack(hists)
timer.toc()
if timer.total_time - last_display > 0.1:
last_display = timer.total_time
print(f"Calib Neighbors {i:08d}: timings {timer.total_time:4.2f}s")
if np.min(np.sum(neighb_hists, axis=1)) > samples_threshold:
break
cumsum = np.cumsum(neighb_hists.T, axis=0)
percentiles = np.sum(cumsum < (keep_ratio * cumsum[hist_n - 1, :]), axis=0)
neighborhood_limits = percentiles
print('\n')
return neighborhood_limits
def ctpn(self, sess, net, image_name):
"""
:param sess: 会话
:param net: 创建的测试网络
:param image_name: 所要测试的单张图片的目录
:return:
"""
timer = Timer()
timer.tic()
# 读取图片
image = cv2.imread(image_name)
shape = image.shape[:2] # 获取高，宽
# resize_im，返回缩放后的图片和相应的缩放比。缩放比定义为 修改后的图/原图
img, scale = TestClass.resize_im(image,
scale=self._cfg.TEST.SCALE,
max_scale=self._cfg.TEST.MAX_SCALE)
# 将图片去均值化
im_orig = img.astype(np.float32, copy=True)
im_orig -= self._cfg.TRAIN.PIXEL_MEANS
# 将缩放和去均值化以后的图片，放入网络进行前向计算，获取分数和对应的文本片段，该片段为映射到最原始图片的坐标
scores, boxes = TestClass.test_ctpn(sess, net, im_orig, scale)
# 此处调用了一个文本检测器
textdetector = TextDetector(self._cfg)
"""
输入参数分别为：
N×4矩阵，每行为一个已经映射回最初的图片的文字片段坐标
N维向量，对应的分数
两维向量，分别为最原始图片的高宽
返回：
一个N×9的矩阵，表示N个拼接以后的完整的文本框。每一行，前八个元素一次是左上，右上，左下，右下的坐标，最后一个元素是文本框的分数
"""
boxes = textdetector.detect(boxes, scores, shape)
self.draw_boxes(image, image_name, boxes, scale)
timer.toc()
print(('Detection took {:.3f}s for '
'{:d} object proposals').format(timer.total_time,
boxes.shape[0]))
def extract_features_batch(model, config, source_path, target_path, voxel_size,
device):
folders = get_folder_list(source_path)
assert len(
folders) > 0, f"Could not find 3DMatch folders under {source_path}"
logging.info(folders)
list_file = os.path.join(target_path, "list.txt")
f = open(list_file, "w")
timer, tmeter = Timer(), AverageMeter()
num_feat = 0
model.eval()
for fo in folders:
if 'evaluation' in fo:
continue
files = get_file_list(fo, ".ply")
fo_base = os.path.basename(fo)
f.write("%s %d\n" % (fo_base, len(files)))
for i, fi in enumerate(files):
# Extract features from a file
pcd = o3d.io.read_point_cloud(fi)
save_fn = "%s_%03d" % (fo_base, i)
if i % 100 == 0:
logging.info(f"{i} / {len(files)}: {save_fn}")
timer.tic()
xyz_down, feature = extract_features(model,
xyz=np.array(pcd.points),
rgb=None,
normal=None,
voxel_size=voxel_size,
device=device,
skip_check=True)
t = timer.toc()
if i > 0:
tmeter.update(t)
num_feat += len(xyz_down)
np.savez_compressed(os.path.join(target_path, save_fn),
points=np.array(pcd.points),
xyz=xyz_down,
feature=feature.detach().cpu().numpy())
if i % 20 == 0 and i > 0:
# 最后一项算的是每个点的特征提取时间
logging.info(
f'Average time: {tmeter.avg}, FPS: {num_feat / tmeter.sum}, time / feat: {tmeter.sum / num_feat}, '
)
f.close()
class CTPNTrainer(BaseTrain):
def __init__(self, sess, model, data, logger):
super(CTPNTrainer, self).__init__(sess, model, data, logger)
self.imdb = data.load_imdb('voc_2007_trainval')
self.roidb = data.get_training_roidb(self.imdb)
self.pretrained_model = cfg.PRETRAINED_MODEL if cfg.PRETRAINED_MODEL else None
# print('Computing bounding-box regression targets...')
# if cfg.TRAIN.BBOX_REG:
# self.bbox_means, self.bbox_stds = data.add_bbox_regression_targets(self.roidb)
# print('done')
self.timer = Timer()
def get_train_op(self, loss):
lr = tf.Variable(cfg.TRAIN.LEARNING_RATE, trainable=False)
if cfg.TRAIN.SOLVER == 'Adam':
opt = tf.train.AdamOptimizer(cfg.TRAIN.LEARNING_RATE)
elif cfg.TRAIN.SOLVER == 'RMS':
opt = tf.train.RMSPropOptimizer(cfg.TRAIN.LEARNING_RATE)
else:
momentum = cfg.TRAIN.MOMENTUM
opt = tf.train.MomentumOptimizer(lr, momentum)
if cfg.TRAIN.WITH_CLIP:
tvars = tf.trainable_variables()
grads, _ = tf.clip_by_global_norm(tf.gradients(loss, tvars), 10.0)
train_op = opt.apply_gradients(list(zip(grads, tvars)),
global_step=self.global_step)
else:
train_op = opt.minimize(loss, global_step=self.global_step)
return train_op, lr
def load_model(self, restore):
restore_iter = 0
if self.pretrained_model is not None and not restore:
try:
print(('Loading pretrained model weights from {:s}').format(
self.pretrained_model))
self.model.load_npz(self.pretrained_model, self.sess, True)
except:
raise 'Check your pretrained model {:s}'.format(
self.pretrained_model)
# resuming a trainer
if restore:
ckpt_path = self.model.load_ckpt(self.sess)
stem = os.path.splitext(os.path.basename(ckpt_path))[0]
restore_iter = int(stem.split('_')[-1])
self.sess.run(self.global_step.assign(restore_iter))
return restore_iter
def train(self, max_iters, restore=False):
"""Network training loop."""
data_layer = DataGenerator(self.roidb, self.imdb.nrof_classes,
self.data)
total_loss, model_loss, rpn_cross_entropy, rpn_loss_box = self.model.build_loss(
ohem=cfg.TRAIN.OHEM)
summary_op, log_image, log_image_data, log_image_name = self.logger.init_summary(
rpn_reg_loss=rpn_loss_box,
rpn_cls_loss=rpn_cross_entropy,
model_loss=model_loss,
total_loss=total_loss)
train_op, lr = self.get_train_op(total_loss)
# intialize variables
self.sess.run(tf.global_variables_initializer())
restore_iter = self.load_model(restore)
fetch_list = [
total_loss, model_loss, rpn_cross_entropy, rpn_loss_box,
summary_op, train_op
]
print(restore_iter, max_iters)
for _iter in range(restore_iter, max_iters, cfg.TRAIN.EPOCH_SIZE):
losses = self.train_epoch(_iter, lr, data_layer, fetch_list)
print('iter: %d / %d, total loss: %.4f, model loss: %.4f, rpn_loss_cls: %.4f, rpn_loss_box: %.4f, lr: %f' % \
(_iter+cfg.TRAIN.EPOCH_SIZE, max_iters, losses[0], losses[1], losses[2], losses[3], losses[5].eval()))
self.logger.summarize(losses[4], self.global_step.eval())
self.save(_iter + cfg.TRAIN.EPOCH_SIZE)
def train_epoch(self, tm_iter, lr, data_layer, fetch_list):
loop = tqdm(range(cfg.TRAIN.EPOCH_SIZE))
for _iter in loop:
tm_iter += _iter
if tm_iter != 0 and tm_iter % cfg.TRAIN.STEPSIZE == 0:
self.sess.run(tf.assign(lr, lr.eval() * cfg.TRAIN.GAMMA))
self.timer.tic()
total_loss_val, model_loss_val, rpn_loss_cls_val, rpn_loss_box_val, summary_str, _ = \
self.train_step(data_layer,fetch_list)
_diff_time = self.timer.toc(average=False)
print('speed: {:.3f}s / iter'.format(_diff_time))
return total_loss_val, model_loss_val, rpn_loss_cls_val, rpn_loss_box_val, summary_str, lr
def train_step(self, data_layer, fetch_list):
blobs = data_layer.forward()
feed_dict = {
self.model.data: blobs['data'],
self.model.im_info: blobs['im_info'],
self.model.keep_prob: 0.5,
self.model.gt_boxes: blobs['gt_boxes'],
self.model.gt_ishard: blobs['gt_ishard'],
self.model.dontcare_areas: blobs['dontcare_areas']
}
return self.sess.run(fetches=fetch_list, feed_dict=feed_dict)
def _valid_epoch(self):
# Change the network to evaluation mode
self.model.eval()
self.val_data_loader.dataset.reset_seed(0)
num_data = 0
hit_ratio_meter, feat_match_ratio, loss_meter, rte_meter, rre_meter = AverageMeter(
), AverageMeter(), AverageMeter(), AverageMeter(), AverageMeter()
data_timer, feat_timer, matching_timer = Timer(), Timer(), Timer()
tot_num_data = len(self.val_data_loader.dataset)
if self.val_max_iter > 0:
tot_num_data = min(self.val_max_iter, tot_num_data)
data_loader_iter = self.val_data_loader.__iter__()
for batch_idx in range(tot_num_data):
data_timer.tic()
input_dict = data_loader_iter.next()
data_timer.toc()
# pairs consist of (xyz1 index, xyz0 index)
feat_timer.tic()
sinput0 = ME.SparseTensor(
input_dict['sinput0_F'], coords=input_dict['sinput0_C']).to(self.device)
F0 = self.model(sinput0).F
sinput1 = ME.SparseTensor(
input_dict['sinput1_F'], coords=input_dict['sinput1_C']).to(self.device)
F1 = self.model(sinput1).F
feat_timer.toc()
matching_timer.tic()
xyz0, xyz1, T_gt = input_dict['pcd0'], input_dict['pcd1'], input_dict['T_gt']
xyz0_corr, xyz1_corr = self.find_corr(xyz0, xyz1, F0, F1, subsample_size=5000)
T_est = te.est_quad_linear_robust(xyz0_corr, xyz1_corr)
loss = corr_dist(T_est, T_gt, xyz0, xyz1, weight=None)
loss_meter.update(loss)
rte = np.linalg.norm(T_est[:3, 3] - T_gt[:3, 3])
rte_meter.update(rte)
rre = np.arccos((np.trace(T_est[:3, :3].t() @ T_gt[:3, :3]) - 1) / 2)
if not np.isnan(rre):
rre_meter.update(rre)
hit_ratio = self.evaluate_hit_ratio(
xyz0_corr, xyz1_corr, T_gt, thresh=self.config.hit_ratio_thresh)
hit_ratio_meter.update(hit_ratio)
feat_match_ratio.update(hit_ratio > 0.05)
matching_timer.toc()
num_data += 1
torch.cuda.empty_cache()
if batch_idx % 100 == 0 and batch_idx > 0:
logging.info(' '.join([
f"Validation iter {num_data} / {tot_num_data} : Data Loading Time: {data_timer.avg:.3f},",
f"Feature Extraction Time: {feat_timer.avg:.3f}, Matching Time: {matching_timer.avg:.3f},",
f"Loss: {loss_meter.avg:.3f}, RTE: {rte_meter.avg:.3f}, RRE: {rre_meter.avg:.3f},",
f"Hit Ratio: {hit_ratio_meter.avg:.3f}, Feat Match Ratio: {feat_match_ratio.avg:.3f}"
]))
data_timer.reset()
logging.info(' '.join([
f"Final Loss: {loss_meter.avg:.3f}, RTE: {rte_meter.avg:.3f}, RRE: {rre_meter.avg:.3f},",
f"Hit Ratio: {hit_ratio_meter.avg:.3f}, Feat Match Ratio: {feat_match_ratio.avg:.3f}"
]))
return {
"loss": loss_meter.avg,
"rre": rre_meter.avg,
"rte": rte_meter.avg,
'feat_match_ratio': feat_match_ratio.avg,
'hit_ratio': hit_ratio_meter.avg
}
def _valid_epoch(self, data_loader_iter):
# Change the network to evaluation mode
self.model.eval()
num_data = 0
hit_ratio_meter, reciprocity_ratio_meter = AverageMeter(
), AverageMeter()
reciprocity_hit_ratio_meter = AverageMeter()
data_timer, feat_timer = Timer(), Timer()
tot_num_data = len(self.val_data_loader.dataset)
if self.val_max_iter > 0:
tot_num_data = min(self.val_max_iter, tot_num_data)
for curr_iter in range(tot_num_data):
data_timer.tic()
input_dict = self.get_data(data_loader_iter)
data_timer.toc()
# pairs consist of (xyz1 index, xyz0 index)
feat_timer.tic()
with torch.no_grad():
F0 = self.model(input_dict['img0'].to(self.device))
F1 = self.model(input_dict['img1'].to(self.device))
feat_timer.toc()
# Test self.num_pos_per_batch * self.batch_size features only.
_, _, H0, W0 = F0.shape
_, _, H1, W1 = F1.shape
for batch_idx, pair in enumerate(input_dict['pairs']):
N = len(pair)
sel = np.random.choice(N,
min(N, self.config.num_pos_per_batch),
replace=False)
curr_pair = pair[sel]
w0, h0, w1, h1 = torch.floor(curr_pair.t() /
self.out_tensor_stride).long()
feats0 = F0[batch_idx, :, h0, w0]
nn_inds1 = find_nn_gpu(feats0,
F1[batch_idx, :].view(F1.shape[1], -1),
nn_max_n=self.config.nn_max_n,
transposed=True)
# Convert the index to coordinate: BxCxHxW
xs1 = nn_inds1 % W1
ys1 = nn_inds1 // W1
# Test reciprocity
nn_inds0 = find_nn_gpu(F1[batch_idx, :, ys1, xs1],
F0[batch_idx, :].view(F0.shape[1], -1),
nn_max_n=self.config.nn_max_n,
transposed=True)
# Convert the index to coordinate: BxCxHxW
xs0 = nn_inds0 % W0
ys0 = nn_inds0 // W0
dist_sq = (w1 - xs1)**2 + (h1 - ys1)**2
is_correct = dist_sq < (self.config.ucn_inlier_threshold_pixel
/ self.out_tensor_stride)**2
hit_ratio_meter.update(is_correct.sum().item() /
len(is_correct))
# Recipocity test result
dist_sq_nn = (w0 - xs0)**2 + (h0 - ys0)**2
mask = dist_sq_nn < (self.config.ucn_inlier_threshold_pixel /
self.out_tensor_stride)**2
reciprocity_ratio_meter.update(mask.sum().item() /
float(len(mask)))
reciprocity_hit_ratio_meter.update(
is_correct[mask].sum().item() / (mask.sum().item() + eps))
torch.cuda.empty_cache()
# visualize_image_correspondence(input_dict['img0'][batch_idx, 0].numpy() + 0.5,
# input_dict['img1'][batch_idx, 0].numpy() + 0.5,
# F0[batch_idx], F1[batch_idx], curr_iter,
# self.config)
num_data += 1
if num_data % 100 == 0:
logging.info(', '.join([
f"Validation iter {num_data} / {tot_num_data} : Data Loading Time: {data_timer.avg:.3f}",
f"Feature Extraction Time: {feat_timer.avg:.3f}, Hit Ratio: {hit_ratio_meter.avg}",
f"Reciprocity Ratio: {reciprocity_ratio_meter.avg}, Reci Filtered Hit Ratio: {reciprocity_hit_ratio_meter.avg}"
]))
data_timer.reset()
logging.info(', '.join([
f"Validation : Data Loading Time: {data_timer.avg:.3f}",
f"Feature Extraction Time: {feat_timer.avg:.3f}, Hit Ratio: {hit_ratio_meter.avg}",
f"Reciprocity Ratio: {reciprocity_ratio_meter.avg}, Reci Filtered Hit Ratio: {reciprocity_hit_ratio_meter.avg}"
]))
return {
'hit_ratio': hit_ratio_meter.avg,
'reciprocity_ratio': reciprocity_ratio_meter.avg,
'reciprocity_hit_ratio': reciprocity_hit_ratio_meter.avg,
}
def main(config):
test_loader = make_data_loader(
config, config.test_phase, 1, num_threads=config.test_num_workers, shuffle=True)
num_feats = 1
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
Model = load_model(config.model)
model = Model(
num_feats,
config.model_n_out,
bn_momentum=config.bn_momentum,
conv1_kernel_size=config.conv1_kernel_size,
normalize_feature=config.normalize_feature)
checkpoint = torch.load(config.save_dir + '/checkpoint.pth')
model.load_state_dict(checkpoint['state_dict'])
model = model.to(device)
model.eval()
success_meter, rte_meter, rre_meter = AverageMeter(), AverageMeter(), AverageMeter()
data_timer, feat_timer, reg_timer = Timer(), Timer(), Timer()
test_iter = test_loader.__iter__()
N = len(test_iter)
n_gpu_failures = 0
# downsample_voxel_size = 2 * config.voxel_size
for i in range(len(test_iter)):
data_timer.tic()
try:
data_dict = test_iter.next()
except ValueError:
n_gpu_failures += 1
logging.info(f"# Erroneous GPU Pair {n_gpu_failures}")
continue
data_timer.toc()
xyz0, xyz1 = data_dict['pcd0'], data_dict['pcd1']
T_gth = data_dict['T_gt']
xyz0np, xyz1np = xyz0.numpy(), xyz1.numpy()
pcd0 = make_open3d_point_cloud(xyz0np)
pcd1 = make_open3d_point_cloud(xyz1np)
with torch.no_grad():
feat_timer.tic()
sinput0 = ME.SparseTensor(
data_dict['sinput0_F'].to(device), coordinates=data_dict['sinput0_C'].to(device))
F0 = model(sinput0).F.detach()
sinput1 = ME.SparseTensor(
data_dict['sinput1_F'].to(device), coordinates=data_dict['sinput1_C'].to(device))
F1 = model(sinput1).F.detach()
feat_timer.toc()
feat0 = make_open3d_feature(F0, 32, F0.shape[0])
feat1 = make_open3d_feature(F1, 32, F1.shape[0])
reg_timer.tic()
distance_threshold = config.voxel_size * 1.0
ransac_result = o3d.registration.registration_ransac_based_on_feature_matching(
pcd0, pcd1, feat0, feat1, distance_threshold,
o3d.registration.TransformationEstimationPointToPoint(False), 4, [
o3d.registration.CorrespondenceCheckerBasedOnEdgeLength(0.9),
o3d.registration.CorrespondenceCheckerBasedOnDistance(
distance_threshold)
], o3d.registration.RANSACConvergenceCriteria(4000000, 10000))
T_ransac = torch.from_numpy(
ransac_result.transformation.astype(np.float32))
reg_timer.toc()
# Translation error
rte = np.linalg.norm(T_ransac[:3, 3] - T_gth[:3, 3])
rre = np.arccos((np.trace(T_ransac[:3, :3].t() @ T_gth[:3, :3]) - 1) / 2)
# Check if the ransac was successful. successful if rte < 2m and rre < 5◦
# http://openaccess.thecvf.com/content_ECCV_2018/papers/Zi_Jian_Yew_3DFeat-Net_Weakly_Supervised_ECCV_2018_paper.pdf
if rte < 2:
rte_meter.update(rte)
if not np.isnan(rre) and rre < np.pi / 180 * 5:
rre_meter.update(rre)
if rte < 2 and not np.isnan(rre) and rre < np.pi / 180 * 5:
success_meter.update(1)
else:
success_meter.update(0)
logging.info(f"Failed with RTE: {rte}, RRE: {rre}")
if i % 10 == 0:
logging.info(
f"{i} / {N}: Data time: {data_timer.avg}, Feat time: {feat_timer.avg}," +
f" Reg time: {reg_timer.avg}, RTE: {rte_meter.avg}," +
f" RRE: {rre_meter.avg}, Success: {success_meter.sum} / {success_meter.count}"
+ f" ({success_meter.avg * 100} %)")
data_timer.reset()
feat_timer.reset()
reg_timer.reset()
logging.info(
f"RTE: {rte_meter.avg}, var: {rte_meter.var}," +
f" RRE: {rre_meter.avg}, var: {rre_meter.var}, Success: {success_meter.sum} " +
f"/ {success_meter.count} ({success_meter.avg * 100} %)")
def _valid_epoch(self):
# Change the network to evaluation mode
self.model.eval()
self.val_data_loader.dataset.reset_seed(0)
num_data = 0
hit_ratio_meter, feat_match_ratio, loss_meter, rte_meter, rre_meter = AverageMeter(
), AverageMeter(), AverageMeter(), AverageMeter(), AverageMeter()
data_timer, feat_timer, matching_timer = Timer(), Timer(), Timer()
tot_num_data = len(self.val_data_loader.dataset)
if self.val_max_iter > 0:
tot_num_data = min(self.val_max_iter, tot_num_data)
data_loader_iter = self.val_data_loader.__iter__()
for batch_idx in range(tot_num_data):
data_timer.tic()
input_dict = data_loader_iter.next()
data_timer.toc()
# pairs consist of (xyz1 index, xyz0 index)
feat_timer.tic()
coords = input_dict['sinput0_C'].to(self.device)
sinput0 = ME.SparseTensor(
input_dict['sinput0_F'].to(self.device),
coordinates=input_dict['sinput0_C'].to(self.device).type(torch.float))
F0 = self.model(sinput0).F
sinput1 = ME.SparseTensor(
input_dict['sinput1_F'].to(self.device),
coordinates=input_dict['sinput1_C'].to(self.device).type(torch.float))
F1 = self.model(sinput1).F
feat_timer.toc()
matching_timer.tic()
xyz0, xyz1, T_gt = input_dict['pcd0'], input_dict['pcd1'], input_dict['T_gt']
xyz0_corr, xyz1_corr = self.find_corr(xyz0, xyz1, F0, F1, subsample_size=5000)
if False:
from sklearn.decomposition import PCA
import open3d as o3d
pc0 = o3d.geometry.PointCloud()
pc0.points = o3d.utility.Vector3dVector(xyz0.numpy())
pca = PCA(n_components=3)
colors = pca.fit_transform(torch.cat((F0, F1), axis=0).cpu().numpy())
colors -= colors.min()
colors /= colors.max()
pc0.colors = o3d.utility.Vector3dVector(colors[0:F0.shape[0]])
o3d.io.write_point_cloud("pc0.ply", pc0)
pc0.transform(T_gt.numpy())
o3d.io.write_point_cloud("pc0_trans.ply", pc0)
pc1 = o3d.geometry.PointCloud()
pc1.points = o3d.utility.Vector3dVector(xyz1.numpy())
pc1.colors = o3d.utility.Vector3dVector(colors[F0.shape[0]:])
o3d.io.write_point_cloud("pc1.ply", pc1)
ind_0 = input_dict['correspondences'][:, 0].type(torch.long)
ind_1 = input_dict['correspondences'][:, 1].type(torch.long)
pc1.points = o3d.utility.Vector3dVector(xyz1[ind_1].numpy())
pc1.colors = o3d.utility.Vector3dVector(
colors[F0.shape[0]:][ind_1])
o3d.io.write_point_cloud("pc1_corr.ply", pc1)
pc0.points = o3d.utility.Vector3dVector(xyz0[ind_0].numpy())
pc0.colors = o3d.utility.Vector3dVector(colors[:F0.shape[0]][ind_0])
pc0.transform(T_gt.numpy())
o3d.io.write_point_cloud("pc0_trans_corr.ply", pc0)
import pdb
pdb.set_trace()
#pc0.points = o3d.utility.Vector3dVector(xyz0_corr.numpy())
# pc0.transform(T_gt.numpy())
#o3d.io.write_point_cloud("xyz0_corr_trans.ply" , pc0)
#
#pc0.points = o3d.utility.Vector3dVector(xyz1_corr.numpy())
#o3d.io.write_point_cloud("xyz1_corr_trans.ply" , pc0)
T_est = te.est_quad_linear_robust(xyz0_corr, xyz1_corr)
loss = corr_dist(T_est, T_gt, xyz0, xyz1, weight=None)
loss_meter.update(loss)
rte = np.linalg.norm(T_est[:3, 3] - T_gt[:3, 3])
rte_meter.update(rte)
rre = np.arccos((np.trace(T_est[:3, :3].t() @ T_gt[:3, :3]) - 1) / 2)
if not np.isnan(rre):
rre_meter.update(rre)
hit_ratio = self.evaluate_hit_ratio(xyz0_corr, xyz1_corr, T_gt, thresh=self.config.hit_ratio_thresh)
hit_ratio_meter.update(hit_ratio)
feat_match_ratio.update(hit_ratio > 0.05)
matching_timer.toc()
num_data += 1
torch.cuda.empty_cache()
if batch_idx % 100 == 0 and batch_idx > 0:
logging.info(' '.join([
f"Validation iter {num_data} / {tot_num_data} : Data Loading Time: {data_timer.avg:.3f},",
f"Feature Extraction Time: {feat_timer.avg:.3f}, Matching Time: {matching_timer.avg:.3f},",
f"Loss: {loss_meter.avg:.3f}, RTE: {rte_meter.avg:.3f}, RRE: {rre_meter.avg:.3f},",
f"Hit Ratio: {hit_ratio_meter.avg:.3f}, Feat Match Ratio: {feat_match_ratio.avg:.3f}"
]))
data_timer.reset()
logging.info(' '.join([
f"Final Loss: {loss_meter.avg:.3f}, RTE: {rte_meter.avg:.3f}, RRE: {rre_meter.avg:.3f},",
f"Hit Ratio: {hit_ratio_meter.avg:.3f}, Feat Match Ratio: {feat_match_ratio.avg:.3f}"
]))
return {
"loss": loss_meter.avg,
"rre": rre_meter.avg,
"rte": rte_meter.avg,
'feat_match_ratio': feat_match_ratio.avg,
'hit_ratio': hit_ratio_meter.avg
}
def test_net(self, graph):
timer = Timer()
timer.tic()
if os.path.exists(self._cfg.TEST.RESULT_DIR_TXT):
shutil.rmtree(self._cfg.TEST.RESULT_DIR_TXT)
os.makedirs(self._cfg.TEST.RESULT_DIR_TXT)
if os.path.exists(self._cfg.TEST.RESULT_DIR_PIC):
shutil.rmtree(self._cfg.TEST.RESULT_DIR_PIC)
os.makedirs(self._cfg.TEST.RESULT_DIR_PIC)
saver = tf.train.Saver()
# 创建一个Session
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allocator_type = 'BFC'
config.gpu_options.per_process_gpu_memory_fraction = 0.7 # 不能太大，否则报错
sess = tf.Session(config=config, graph=graph)
# 获取一个Saver()实例
# 恢复模型参数
ckpt = tf.train.get_checkpoint_state(self._cfg.COMMON.CKPT)
if ckpt and ckpt.model_checkpoint_path:
print('Restoring from {}...'.format(ckpt.model_checkpoint_path), end=' ')
try:
saver.restore(sess, ckpt.model_checkpoint_path)
except:
raise 'Check your pretrained {:s}'.format(ckpt.model_checkpoint_path)
print('done')
else:
raise 'Check your pretrained {:s}'.format(self._cfg.TEST.RESULT_DIR)
# # TODO 这里需要仔细测试一下
# im_names = glob.glob(os.path.join(self._cfg.TEST.DATA_DIR, '*.png')) + \
# glob.glob(os.path.join(self._cfg.TEST.DATA_DIR, '*.jpg'))
im_names = os.listdir(self._cfg.TEST.DATA_DIR)
assert len(im_names) > 0, "Nothing to test"
i = 0
for im in im_names:
im_name = os.path.join(self._cfg.TEST.DATA_DIR, im)
# print('~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~')
# print(('Testing for image {:s}'.format(im_name)))
try:
self.ctpn(sess, self._net, im_name)
except NoPositiveError:
print("Warning!!, get no region of interest in picture {}".format(im))
continue
except:
print("the pic {} may has problems".format(im))
continue
i += 1
if i % 10 == 0:
timer.toc()
print('Detection took {:.3f}s for 10 pic'.format(timer.total_time))
# 最后关闭session
sess.close()
def _train_epoch(self, epoch):
gc.collect()
self.model.train()
# Epoch starts from 1
self.data_loader.sampler.set_epoch(epoch)
data_timer, total_timer = Timer(), Timer()
start_iter = (epoch - 1) * len(self.data_loader)
data_timer.tic()
total_timer.tic()
for curr_iter, input_dict in enumerate(self.data_loader):
data_timer.toc()
self.optimizer.zero_grad()
batch_pos_loss, batch_neg_loss, batch_loss = 0, 0, 0
sinput0 = ME.SparseTensor(input_dict['sinput0_F'].to(self.device),
coordinates=input_dict['sinput0_C'].to(
self.device))
F0 = self.model(sinput0).F
sinput1 = ME.SparseTensor(input_dict['sinput1_F'].to(self.device),
coordinates=input_dict['sinput1_C'].to(
self.device))
F1 = self.model(sinput1).F
pos_pairs = input_dict['correspondences']
pos_loss, neg_loss = self.contrastive_hardest_negative_loss(
input_dict['pcd0_rot'],
input_dict['pcd1'],
F0,
F1,
pos_pairs,
num_pos=self.config.num_pos_per_batch * self.config.batch_size,
num_hn_samples=self.config.num_hn_samples_per_batch *
self.config.batch_size,
matching_search_voxel_size=self.config.voxel_size *
self.config.positive_pair_search_voxel_size_multiplier)
loss = pos_loss + self.neg_weight * neg_loss
loss.backward()
batch_loss += loss.item()
batch_pos_loss += pos_loss.item()
batch_neg_loss += neg_loss.item()
self.sum_gradients()
self.optimizer.step()
torch.cuda.empty_cache()
total_timer.toc()
if curr_iter % self.config.stat_freq == 0:
report = torch.tensor(
[1.0, batch_loss, batch_pos_loss, batch_neg_loss],
device=torch.cuda.current_device())
dist.all_reduce(report, op=dist.ReduceOp.SUM)
count = report[0].item()
m_batch_loss = report[1].item() / count
m_batch_pos_loss = report[2].item() / count
m_batch_neg_loss = report[3].item() / count
if self.rank == 0:
self.writer.add_scalar('train/loss', m_batch_loss,
start_iter + curr_iter)
self.writer.add_scalar('train/pos_loss', m_batch_pos_loss,
start_iter + curr_iter)
self.writer.add_scalar('train/neg_loss', m_batch_neg_loss,
start_iter + curr_iter)
logging.info(
"Train Epoch: {} [{}/{}], Current Loss: {:.3e} Pos: {:.3f} Neg: {:.3f}"
.format(epoch, curr_iter, len(
self.data_loader), m_batch_loss, m_batch_pos_loss,
m_batch_neg_loss) +
"\tData time: {:.4f}, Train time: {:.4f}, Iter time: {:.4f}"
.format(data_timer.avg, total_timer.avg -
data_timer.avg, total_timer.avg))
total_timer.reset()
total_timer.tic()
data_timer.tic()
def main(config):
test_loader = make_data_loader(config,
config.test_phase,
1,
num_threads=config.test_num_thread,
shuffle=False)
num_feats = 1
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
Model = load_model(config.model)
model = Model(num_feats,
config.model_n_out,
bn_momentum=config.bn_momentum,
conv1_kernel_size=config.conv1_kernel_size,
normalize_feature=config.normalize_feature)
checkpoint = torch.load(config.save_dir + '/checkpoint.pth')
model.load_state_dict(checkpoint['state_dict'])
model = model.to(device)
model.eval()
success_meter, rte_meter, rre_meter = AverageMeter(), AverageMeter(
), AverageMeter()
data_timer, feat_timer, reg_timer = Timer(), Timer(), Timer()
test_iter = test_loader.__iter__()
N = len(test_iter)
n_gpu_failures = 0
# downsample_voxel_size = 2 * config.voxel_size
list_results_to_save = []
for i in range(len(test_iter)):
data_timer.tic()
try:
data_dict = test_iter.next()
except ValueError:
n_gpu_failures += 1
logging.info(f"# Erroneous GPU Pair {n_gpu_failures}")
continue
data_timer.toc()
xyz0, xyz1 = data_dict['pcd0'], data_dict['pcd1']
T_gth = data_dict['T_gt']
xyz0np, xyz1np = xyz0.numpy(), xyz1.numpy()
#import pdb
# pdb.set_trace()
pcd0 = make_open3d_point_cloud(xyz0np)
pcd1 = make_open3d_point_cloud(xyz1np)
with torch.no_grad():
feat_timer.tic()
sinput0 = ME.SparseTensor(
data_dict['sinput0_F'].to(device),
coordinates=data_dict['sinput0_C'].to(device))
F0 = model(sinput0).F.detach()
sinput1 = ME.SparseTensor(
data_dict['sinput1_F'].to(device),
coordinates=data_dict['sinput1_C'].to(device))
F1 = model(sinput1).F.detach()
feat_timer.toc()
# saving files to pkl
print(i)
dict_sample = {
"pts_source": xyz0np,
"feat_source": F0.cpu().detach().numpy(),
"pts_target": xyz1np,
"feat_target": F1.cpu().detach().numpy()
}
list_results_to_save.append(dict_sample)
import pickle
path_results_to_save = "fcgf.results.pkl"
print('Saving results to ', path_results_to_save)
pickle.dump(list_results_to_save, open(path_results_to_save, 'wb'))
print('Saved!')
import pdb
pdb.set_trace()
def dump_correspondences(config):
# load model
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
checkpoint = torch.load(config.weights)
model = ResUNetBN2D2(1, 64, normalize_feature=True)
model.load_state_dict(checkpoint['state_dict'])
model.eval()
model = model.to(device)
print("load model")
# load dataset
source = config.source
with h5py.File(config.target, 'r+') as ofp:
new_data = {}
keys = ['ucn_coords', 'ucn_n_coords']
for k in keys:
if k in ofp.keys():
new_data[k] = ofp[k]
else:
new_data[k] = ofp.create_group(k)
len_dset = len(ofp['coords'])
keys = ofp['ucn_coords'].keys()
print("len dataset : ", len_dset)
matching_timer, write_timer = Timer(), Timer()
# extract correspondences
for i in range(len_dset):
# skip existing pair
if str(i) in keys:
continue
_coords = ofp['coords'][str(i)]
img_path0 = _coords.attrs['img0']
img_path1 = _coords.attrs['img1']
img_idx0 = int(_coords.attrs['idx0']) + 1
img_idx1 = int(_coords.attrs['idx1']) + 1
calib_path0 = "/".join(img_path0.split("/")[:-2])
calib_path0 += f"/calibration/calibration_{img_idx0:06d}.h5"
calib_path1 = "/".join(img_path1.split("/")[:-2])
calib_path1 += f"/calibration/calibration_{img_idx1:06d}.h5"
img0 = prep_image(osp.join(source, img_path0))
img1 = prep_image(osp.join(source, img_path1))
F0 = model(to_normalized_torch(img0, device))
F1 = model(to_normalized_torch(img1, device))
args = (img0, img1, calib_path0, calib_path1, F0, F1, i, len_dset,
source)
matching_timer.tic()
coords, n_coords = dump_correspondence_single(args)
matching_timer.toc()
write_timer.tic()
coords_data = new_data['ucn_coords'].create_dataset(
str(i), coords.shape, dtype=np.float32)
coords_data[:] = coords.astype(np.float32)
n_coords_data = new_data['ucn_n_coords'].create_dataset(
str(i), n_coords.shape, dtype=np.float32)
n_coords_data[:] = n_coords.astype(np.float32)
write_timer.toc()
print(
f"[{i}/{len_dset}] save {coords.shape} coordinate, matching {matching_timer.avg:.3f}, write {write_timer.avg:.3f}"
)
