def __call__(self, scan):
prob = random.random()
if prob > self.augm_prob:
return scan
else:
diff = random.uniform(-self.max_height_diff, self.max_height_diff)
points = scan.points + np.array([diff, 0, 0])
theta = np.arcsin(scan.points[:, 2] /
np.linalg.norm(scan.points, axis=1))
fov_up = np.min([
np.max(theta) * 180 / np.pi,
scan.proj_fov_up + self.max_angle_diff
])
fov_down = np.max([
np.min(theta) * 180 / np.pi,
scan.proj_fov_down - self.max_angle_diff
])
new_laserscan = SemLaserScan(self.color_map,
project=scan.project,
H=scan.proj_H,
W=scan.proj_W,
fov_up=fov_up,
fov_down=fov_down)
new_laserscan.set_points(points, scan.remissions)
return new_laserscan
def __getitem__(self, index):
frame_num = self.frame_num
proj_colors_list = []
path_seq_list = []
path_name_list = []
for idx in range(index * frame_num, index * frame_num + frame_num):
# get item in tensor shape
scan_file = self.scan_files[idx]
if self.gt:
label_file = self.label_files[idx]
# open a semantic laserscan
if self.gt:
scan = SemLaserScan(self.color_map,
project=True,
H=self.sensor_img_H,
W=1024,
fov_up=self.sensor_fov_up,
fov_down=self.sensor_fov_down)
else:
scan = LaserScan(project=True,
H=self.sensor_img_H,
W=self.sensor_img_W,
fov_up=self.sensor_fov_up,
fov_down=self.sensor_fov_down)
# open and obtain scan
scan.open_scan(scan_file)
if self.gt:
scan.open_label(label_file)
# map unused classes to used classes (also for projection)
scan.sem_label = self.map(scan.sem_label, self.learning_map)
mask = scan.proj_idx >= 0
scan.proj_sem_color[mask] = scan.sem_color_lut[scan.sem_label[
scan.proj_idx[mask]]]
if self.gt:
proj_colors = torch.from_numpy(scan.proj_sem_color).clone()
else:
proj_colors = []
# get name and sequence
path_norm = os.path.normpath(scan_file)
path_split = path_norm.split(os.sep)
path_seq = path_split[-3]
path_name = path_split[-1].replace(".bin", ".label")
if self.gt:
proj_colors_list.append(proj_colors.unsqueeze(0))
path_seq_list.append(path_seq)
path_name_list.append(path_name)
if self.gt:
proj_colors_seq = torch.cat(proj_colors_list, dim=0)
else:
proj_colors_seq = []
# return
return proj_colors_seq, path_seq_list, path_name_list
def __call__(self, scan):
v_prob = random.random()
h_prob = random.random()
if v_prob > self.augm_prob and h_prob > self.augm_prob:
return scan
else:
v_flip = True
h_flip = True
if v_prob > self.augm_prob:
v_flip = False
if h_flip > self.augm_prob:
h_flip = False
new_laserscan = SemLaserScan(self.color_map,
project=scan.project,
H=scan.proj_H,
W=scan.proj_W,
fov_up=scan.proj_fov_up,
fov_down=scan.proj_fov_down,
h_flip=h_flip,
v_flip=v_flip)
new_laserscan.set_points(scan.points, scan.remissions)
return new_laserscan
def eval(test_sequences, splits, pred):
# get scan paths
scan_names = []
for sequence in test_sequences:
sequence = '{0:02d}'.format(int(sequence))
scan_paths = os.path.join(FLAGS.dataset, "sequences", str(sequence),
"velodyne")
# populate the scan names
seq_scan_names = [
os.path.join(dp, f)
for dp, dn, fn in os.walk(os.path.expanduser(scan_paths))
for f in fn if ".bin" in f
]
seq_scan_names.sort()
scan_names.extend(seq_scan_names)
# print(scan_names)
# get label paths
label_names = []
for sequence in test_sequences:
sequence = '{0:02d}'.format(int(sequence))
label_paths = os.path.join(FLAGS.dataset, "sequences", str(sequence),
"labels")
# populate the label names
seq_label_names = [
os.path.join(dp, f)
for dp, dn, fn in os.walk(os.path.expanduser(label_paths))
for f in fn if ".label" in f
]
seq_label_names.sort()
label_names.extend(seq_label_names)
# print(label_names)
# get predictions paths
pred_names = []
for sequence in test_sequences:
sequence = '{0:02d}'.format(int(sequence))
pred_paths = os.path.join(FLAGS.predictions, "sequences", sequence,
"predictions")
# populate the label names
seq_pred_names = [
os.path.join(dp, f)
for dp, dn, fn in os.walk(os.path.expanduser(pred_paths))
for f in fn if ".label" in f
]
seq_pred_names.sort()
pred_names.extend(seq_pred_names)
# print(pred_names)
# check that I have the same number of files
print("labels: ", len(label_names))
# for label_name in label_names:
# print(label_name)
print("predictions: ", len(pred_names))
# for pred_name in pred_names:
# print(pred_name)
assert (len(label_names) == len(scan_names)
and len(label_names) == len(pred_names))
print("Evaluating sequences: ")
# open each file, get the tensor, and make the iou comparison
for scan_file, label_file, pred_file in zip(scan_names, label_names,
pred_names):
print("evaluating label ", label_file, "with", pred_file)
# open label
label = SemLaserScan(project=False)
label.open_scan(scan_file)
label.open_label(label_file)
u_label_sem = remap_lut[label.sem_label] # remap to xentropy format
if FLAGS.limit is not None:
u_label_sem = u_label_sem[:FLAGS.limit]
# open prediction
pred = SemLaserScan(project=False)
pred.open_scan(scan_file)
pred.open_label(pred_file)
u_pred_sem = remap_lut[pred.sem_label] # remap to xentropy format
if FLAGS.limit is not None:
u_pred_sem = u_pred_sem[:FLAGS.limit]
# add single scan to evaluation
evaluator.addBatch(u_pred_sem, u_label_sem)
# when I am done, print the evaluation
m_accuracy = evaluator.getacc()
m_jaccard, class_jaccard = evaluator.getIoU()
print('{split} set:\n'
'Acc avg {m_accuracy:.3f}\n'
'IoU avg {m_jaccard:.3f}'.format(split=splits,
m_accuracy=m_accuracy,
m_jaccard=m_jaccard))
save_to_log(
FLAGS.predictions, 'pred.txt', '{split} set:\n'
'Acc avg {m_accuracy:.3f}\n'
'IoU avg {m_jaccard:.3f}'.format(split=splits,
m_accuracy=m_accuracy,
m_jaccard=m_jaccard))
# print also classwise
for i, jacc in enumerate(class_jaccard):
if i not in ignore:
print('IoU class {i:} [{class_str:}] = {jacc:.3f}'.format(
i=i, class_str=class_strings[class_inv_remap[i]], jacc=jacc))
save_to_log(
FLAGS.predictions, 'pred.txt',
'IoU class {i:} [{class_str:}] = {jacc:.3f}'.format(
i=i,
class_str=class_strings[class_inv_remap[i]],
jacc=jacc))
# print for spreadsheet
print("*" * 80)
print("below can be copied straight for paper table")
for i, jacc in enumerate(class_jaccard):
if i not in ignore:
sys.stdout.write('{jacc:.3f}'.format(jacc=jacc.item()))
sys.stdout.write(",")
sys.stdout.write('{jacc:.3f}'.format(jacc=m_jaccard.item()))
sys.stdout.write(",")
sys.stdout.write('{acc:.3f}'.format(acc=m_accuracy.item()))
sys.stdout.write('\n')
sys.stdout.flush()
from common.laserscan import LaserScan, SemLaserScan
parser = argparse.ArgumentParser("./visualize.py")
parser.add_argument(
'--config',
'-c',
type=str,
required=False,
default="config/labels/semantic-kitti.yaml",
help='Dataset config file. Defaults to %(default)s',
)
FLAGS, unparsed = parser.parse_known_args()
CFG = yaml.safe_load(open(FLAGS.config, 'r'))
color_dict = CFG["color_map"]
scan = SemLaserScan(color_dict, project=False)
scan_root = r'G:\DataSet\semanticKITTI\dataset\sequences'
gt_label_root = r'G:\DataSet\semanticKITTI\dataset\sequences'
gt_img_root = r'G:\DataSet\semanticKITTI\visualization\groundtruth\sequences'
segv2_label_root = r'G:\DataSet\semanticKITTI\prediction\SqueezeSegV2\Static\sequences'
segv2_img_root = r'G:\DataSet\semanticKITTI\visualization\SqueezeSegV2\Static\sequences'
segv2_sap1_label_root = r'G:\DataSet\semanticKITTI\prediction\SqueezeSegV2\SAP-1\sequences'
segv2_sap1_img_root = r'G:\DataSet\semanticKITTI\visualization\SqueezeSegV2\SAP-1\sequences'
segv2_asap1_label_root = r'G:\DataSet\semanticKITTI\prediction\SqueezeSegV2\ASAP-1\sequences'
segv2_asap1_img_root = r'G:\DataSet\semanticKITTI\visualization\SqueezeSegV2\ASAP-1\sequences'
segv2_asap2_label_root = r'G:\DataSet\semanticKITTI\prediction\SqueezeSegV2\ASAP-2\sequences'
def __getitem__(self, index):
# get item in tensor shape
scan_file = self.scan_files[index]
if self.gt:
label_file = self.label_files[index]
# open a semantic laserscan
if self.gt:
scan = SemLaserScan(self.color_map,
project=True,
H=self.sensor_img_H,
W=self.sensor_img_W,
fov_up=self.sensor_fov_up,
fov_down=self.sensor_fov_down)
else:
scan = LaserScan(project=True,
H=self.sensor_img_H,
W=self.sensor_img_W,
fov_up=self.sensor_fov_up,
fov_down=self.sensor_fov_down)
# open and obtain scan
scan.open_scan(scan_file)
if self.gt:
scan.open_label(label_file)
# map unused classes to used classes (also for projection)
scan.sem_label = self.map(scan.sem_label, self.learning_map)
scan.proj_sem_label = self.map(scan.proj_sem_label,
self.learning_map)
# make a tensor of the uncompressed data (with the max num points)
unproj_n_points = scan.points.shape[0]
unproj_xyz = torch.full((self.max_points, 3), -1.0, dtype=torch.float)
unproj_xyz[:unproj_n_points] = torch.from_numpy(scan.points)
unproj_range = torch.full([self.max_points], -1.0, dtype=torch.float)
unproj_range[:unproj_n_points] = torch.from_numpy(scan.unproj_range)
unproj_remissions = torch.full([self.max_points],
-1.0,
dtype=torch.float)
unproj_remissions[:unproj_n_points] = torch.from_numpy(scan.remissions)
if self.gt:
unproj_labels = torch.full([self.max_points],
-1.0,
dtype=torch.int32)
unproj_labels[:unproj_n_points] = torch.from_numpy(scan.sem_label)
else:
unproj_labels = []
# get points and labels
proj_range = torch.from_numpy(scan.proj_range).clone()
proj_xyz = torch.from_numpy(scan.proj_xyz).clone()
proj_remission = torch.from_numpy(scan.proj_remission).clone()
proj_mask = torch.from_numpy(scan.proj_mask)
if self.gt:
proj_labels = torch.from_numpy(scan.proj_sem_label).clone()
proj_labels = proj_labels * proj_mask
else:
proj_labels = []
proj_x = torch.full([self.max_points], -1, dtype=torch.long)
proj_x[:unproj_n_points] = torch.from_numpy(scan.proj_x)
proj_y = torch.full([self.max_points], -1, dtype=torch.long)
proj_y[:unproj_n_points] = torch.from_numpy(scan.proj_y)
proj = torch.cat([
proj_range.unsqueeze(0).clone(),
proj_xyz.clone().permute(2, 0, 1),
proj_remission.unsqueeze(0).clone()
])
proj = (proj - self.sensor_img_means[:, None, None]
) / self.sensor_img_stds[:, None, None]
proj = proj * proj_mask.float()
# get name and sequence
path_norm = os.path.normpath(scan_file)
path_split = path_norm.split(os.sep)
path_seq = path_split[-3]
path_name = path_split[-1].replace(".bin", ".label")
# print("path_norm: ", path_norm)
# print("path_seq", path_seq)
# print("path_name", path_name)
# return
return proj, proj_mask, proj_labels, unproj_labels, path_seq, path_name, proj_x, proj_y, proj_range, unproj_range, proj_xyz, unproj_xyz, proj_remission, unproj_remissions, unproj_n_points
pred_names.extend(seq_pred_names)
# print(pred_names)
# check that I have the same number of files
# print("labels: ", len(label_names))
# print("predictions: ", len(pred_names))
assert (len(label_names) == len(scan_names)
and len(label_names) == len(pred_names))
print("Evaluating sequences: ")
# open each file, get the tensor, and make the iou comparison
for scan_file, label_file, pred_file in zip(scan_names, label_names,
pred_names):
print("evaluating label ", label_file, "with", pred_file)
# open label
label = SemLaserScan(project=False)
label.open_scan(scan_file)
label.open_label(label_file)
u_label_sem = remap_lut[label.sem_label] # remap to xentropy format
if FLAGS.limit is not None:
u_label_sem = u_label_sem[:FLAGS.limit]
# open prediction
pred = SemLaserScan(project=False)
pred.open_scan(scan_file)
pred.open_label(pred_file)
u_pred_sem = remap_lut[pred.sem_label] # remap to xentropy format
if FLAGS.limit is not None:
u_pred_sem = u_pred_sem[:FLAGS.limit]
# add single scan to evaluation
quit()
# populate the pointclouds
label_names = [os.path.join(dp, f) for dp, dn, fn in os.walk(
os.path.expanduser(label_paths)) for f in fn]
label_names.sort()
# check that there are same amount of labels and scans
if not FLAGS.ignore_safety:
assert(len(label_names) == len(scan_names))
# create a scan
if FLAGS.ignore_semantics:
scan = LaserScan(project=True) # project all opened scans to spheric proj
else:
color_dict = CFG["color_map"]
scan = SemLaserScan(color_dict, project=True)
# create a visualizer
semantics = not FLAGS.ignore_semantics
if not semantics:
label_names = None
vis = LaserScanVis(scan=scan,
scan_names=scan_names,
label_names=label_names,
offset=FLAGS.offset,
semantics=semantics,
instances=False)
# print instructions
print("To navigate:")
print("\tb: back (previous scan)")
# get device
if torch.cuda.is_available():
device = torch.device('cuda')
else:
device = torch.device('cpu')
# define the border mask
bm = borderMask(300, device, FLAGS.border, FLAGS.conn, FLAGS.exclude_class)
# imports for inference part
import cv2
import numpy as np
from common.laserscan import SemLaserScan
# open label and project
scan = SemLaserScan(project=True, max_classes=300)
scan.open_scan(FLAGS.scan)
scan.open_label(FLAGS.label)
# get the things I need
proj_range = torch.from_numpy(scan.proj_range).to(device)
proj_sem_label = torch.from_numpy(scan.proj_sem_label).long().to(device)
proj_sem_color = torch.from_numpy(scan.proj_sem_color).to(device)
# run the border mask
border_mask = bm(proj_sem_label)
# bring to numpy and normalize for showing
proj_range = proj_range.cpu().numpy()
proj_sem_label = proj_sem_label.cpu().numpy()
proj_sem_color = proj_sem_color.cpu().numpy()
def __getitem__(self, index):
# get item in tensor shape
scan_file = self.scan_files[index]
if self.gt:
label_file = self.label_files[index]
# open a semantic laserscan
DA = False
flip_sign = False
rot = False
drop_points = False
if self.transform:
if random.random() > 0.5:
if random.random() > 0.5:
DA = True
if random.random() > 0.5:
flip_sign = True
if random.random() > 0.5:
rot = True
drop_points = random.uniform(0, 0.5)
if self.gt:
scan = SemLaserScan(self.color_map,
project=True,
H=self.sensor_img_H,
W=self.sensor_img_W,
fov_up=self.sensor_fov_up,
fov_down=self.sensor_fov_down,
DA=DA,
flip_sign=flip_sign,
drop_points=drop_points)
else:
scan = LaserScan(project=True,
H=self.sensor_img_H,
W=self.sensor_img_W,
fov_up=self.sensor_fov_up,
fov_down=self.sensor_fov_down,
DA=DA,
rot=rot,
flip_sign=flip_sign,
drop_points=drop_points)
# open and obtain scan
scan.open_scan(scan_file)
if self.gt:
scan.open_label(label_file)
# map unused classes to used classes (also for projection)
scan.sem_label = self.map(scan.sem_label, self.learning_map)
scan.proj_sem_label = self.map(scan.proj_sem_label, self.learning_map)
# get points and labels
proj_range = torch.from_numpy(scan.proj_range).clone()
proj_segment_angle = torch.from_numpy(scan.segment_angle).clone()
proj_xyz = torch.from_numpy(scan.proj_xyz).clone()
proj_remission = torch.from_numpy(scan.proj_remission).clone()
proj_mask = torch.from_numpy(scan.proj_mask)
if self.gt:
proj_labels = torch.from_numpy(scan.proj_sem_label).clone()
proj_labels = proj_labels * proj_mask
else:
proj_labels = []
# return
return proj_range, proj_segment_angle, proj_xyz, proj_remission, proj_mask, proj_labels, scan.points, scan_file, scan.sem_label
label_names.sort()
# check that there are same amount of labels and scans
if not FLAGS.ignore_safety:
assert (len(label_names) == len(scan_names))
# create a scan
if FLAGS.ignore_semantics:
scan = LaserScan(
project=True) # project all opened scans to spheric proj
else:
color_dict = CFG["color_map"]
#scan = SemLaserScan(color_dict, project=True)
scan = SemLaserScan(color_dict,
project=True,
H=64,
W=1024,
fov_up=3,
fov_down=-25)
# create a visualizer
semantics = not FLAGS.ignore_semantics
if not semantics:
label_names = None
vis = LaserScanVis(scan=scan,
scan_names=scan_names,
label_names=label_names,
offset=FLAGS.offset,
semantics=semantics,
instances=False)
# print instructions
def __getitem__(self, index):
# get item in tensor shape
scan_file = self.scan_files[index]
# JLLIU
# print(scan_file)
if self.gt:
label_file = self.label_files[index]
# open a semantic laserscan
DA = False
flip_sign = False
rot = False
drop_points = False
if self.transform:
if random.random() > 0.5:
if random.random() > 0.5:
DA = True
if random.random() > 0.5:
flip_sign = True
if random.random() > 0.5:
rot = True
drop_points = random.uniform(0, 0.5)
if self.gt:
scan = SemLaserScan(self.color_map,
project=True,
H=self.sensor_img_H,
W=self.sensor_img_W,
fov_up=self.sensor_fov_up,
fov_down=self.sensor_fov_down,
DA=DA,
flip_sign=flip_sign,
drop_points=drop_points)
else:
scan = LaserScan(project=True,
H=self.sensor_img_H,
W=self.sensor_img_W,
fov_up=self.sensor_fov_up,
fov_down=self.sensor_fov_down,
DA=DA,
rot=rot,
flip_sign=flip_sign,
drop_points=drop_points)
# open and obtain scan
scan.open_scan(scan_file)
if self.gt:
scan.open_label(label_file)
# map unused classes to used classes (also for projection)
scan.sem_label = self.map(scan.sem_label, self.learning_map)
scan.proj_sem_label = self.map(scan.proj_sem_label,
self.learning_map)
#print("proj_x.shape: ",proj_x.shape)
#print("proj_y.shape: ",proj_y.shape)
# make a tensor of the uncompressed data (with the max num points)
unproj_n_points = scan.points.shape[0]
unproj_xyz = torch.full((self.max_points, 3), -1.0, dtype=torch.float)
unproj_xyz[:unproj_n_points] = torch.from_numpy(scan.points)
unproj_range = torch.full([self.max_points], -1.0, dtype=torch.float)
unproj_range[:unproj_n_points] = torch.from_numpy(scan.unproj_range)
unproj_remissions = torch.full([self.max_points],
-1.0,
dtype=torch.float)
unproj_remissions[:unproj_n_points] = torch.from_numpy(scan.remissions)
if self.gt:
unproj_labels = torch.full([self.max_points],
-1.0,
dtype=torch.int32)
unproj_labels[:unproj_n_points] = torch.from_numpy(scan.sem_label)
else:
unproj_labels = []
# JLLIU:
# 底下的 proj_xyz.shape: torch.Size([64, 2048, 3]) 就是之前在 laserscan.py 看到的那個
# intensity 的 proj_remission[proj_y, proj_x] 也是
"""
seq_now = os.path.normpath(scan_file).split(os.sep)[-3]
file_now = os.path.normpath(scan_file).split(os.sep)[-1].replace(".bin", ".npy")
proj_xyzi = np.zeros(shape=(64,2048,4))
proj_xyzi[:,:,0:3] = scan.proj_xyz
proj_xyzi[:,:,3] = scan.proj_remission
np.save('/home/doggy/SalsaNext/train/tasks/semantic/dataset/xyzi_npy/'+ seq_now + '/' + file_now , proj_xyzi)
print("\nnpy_name: ", file_now)
print("proj_xyz[50,50]: ", scan.proj_xyz[50,50])
print("proj_i[50,50]: ", scan.proj_remission[50,50])
print("proj_xyzi[50,50]: ", proj_xyzi[50,50])
"""
# get points and labels
proj_range = torch.from_numpy(scan.proj_range).clone()
proj_xyz = torch.from_numpy(scan.proj_xyz).clone()
proj_remission = torch.from_numpy(scan.proj_remission).clone()
proj_mask = torch.from_numpy(scan.proj_mask)
if self.gt:
proj_labels = torch.from_numpy(scan.proj_sem_label).clone()
proj_labels = proj_labels * proj_mask
else:
proj_labels = []
proj_x = torch.full([self.max_points], -1, dtype=torch.long)
proj_x[:unproj_n_points] = torch.from_numpy(scan.proj_x)
proj_y = torch.full([self.max_points], -1, dtype=torch.long)
proj_y[:unproj_n_points] = torch.from_numpy(scan.proj_y)
proj = torch.cat([
proj_range.unsqueeze(0).clone(),
proj_xyz.clone().permute(2, 0, 1),
proj_remission.unsqueeze(0).clone()
])
proj = (proj - self.sensor_img_means[:, None, None]
) / self.sensor_img_stds[:, None, None]
proj = proj * proj_mask.float()
# get name and sequence
path_norm = os.path.normpath(scan_file)
path_split = path_norm.split(os.sep)
path_seq = path_split[-3]
path_name = path_split[-1].replace(".bin", ".label")
# return
return proj, proj_mask, proj_labels, unproj_labels, path_seq, path_name, proj_x, proj_y, proj_range, unproj_range, proj_xyz, unproj_xyz, proj_remission, unproj_remissions, unproj_n_points
for dp, dn, fn in os.walk(os.path.expanduser(label_paths))
for f in fn
]
label_names.sort()
# check that there are same amount of labels and scans
if not FLAGS.ignore_safety:
assert (len(label_names) == len(scan_names))
# create a scan
if FLAGS.ignore_semantics:
scan = LaserScan(
project=True) # project all opened scans to spheric proj
else:
color_dict = CFG["color_map"]
scan = SemLaserScan(color_dict, project=True, H=80, W=2048)
# create a visualizer
semantics = not FLAGS.ignore_semantics
if not semantics:
label_names = None
vis = LaserScanVis(scan=scan,
scan_names=scan_names,
label_names=label_names,
offset=FLAGS.offset,
semantics=semantics,
instances=False)
# print instructions
print("To navigate:")
print("\tb: back (previous scan)")
def __getitem__(self, index):
# get item in tensor shape
scan_file = self.scan_files[index]
if self.gt:
label_file = self.label_files[index]
# open a semantic laserscan
if self.gt:
scan = SemLaserScan(self.color_map,
project=True,
H=self.sensor_img_H,
W=self.sensor_img_W,
fov_up=self.sensor_fov_up,
fov_down=self.sensor_fov_down)
else:
scan = LaserScan(project=True,
H=self.sensor_img_H,
W=self.sensor_img_W,
fov_up=self.sensor_fov_up,
fov_down=self.sensor_fov_down)
# open and obtain scan
scan.open_scan(scan_file)
if self.gt:
scan.open_label(label_file)
# map unused classes to used classes (also for projection)
scan.sem_label = self.map(scan.sem_label, self.learning_map)
scan.proj_sem_label = self.map(scan.proj_sem_label,
self.learning_map)
# make a tensor of the uncompressed data (with the max num points)
unproj_n_points = scan.points.shape[0]
unproj_xyz = torch.full((self.max_points, 3), -1.0, dtype=torch.float)
unproj_xyz[:unproj_n_points] = torch.from_numpy(scan.points)
unproj_range = torch.full([self.max_points], -1.0, dtype=torch.float)
unproj_range[:unproj_n_points] = torch.from_numpy(scan.unproj_range)
unproj_remissions = torch.full([self.max_points],
-1.0,
dtype=torch.float)
unproj_remissions[:unproj_n_points] = torch.from_numpy(scan.remissions)
if self.gt:
unproj_labels = torch.full([self.max_points],
-1.0,
dtype=torch.int32)
unproj_labels[:unproj_n_points] = torch.from_numpy(scan.sem_label)
else:
unproj_labels = []
# get points and labels
proj_range = torch.from_numpy(scan.proj_range).clone()
proj_xyz = torch.from_numpy(scan.proj_xyz).clone()
proj_remission = torch.from_numpy(scan.proj_remission).clone()
proj_mask = torch.from_numpy(scan.proj_mask)
if self.gt:
proj_labels = torch.from_numpy(scan.proj_sem_label).clone()
proj_labels = proj_labels * proj_mask
else:
proj_labels = []
proj_x = torch.full([self.max_points], -1, dtype=torch.long)
proj_x[:unproj_n_points] = torch.from_numpy(scan.proj_x)
proj_y = torch.full([self.max_points], -1, dtype=torch.long)
proj_y[:unproj_n_points] = torch.from_numpy(scan.proj_y)
proj = torch.cat([
proj_range.unsqueeze(0).clone(),
proj_xyz.clone().permute(2, 0, 1),
proj_remission.unsqueeze(0).clone()
])
proj_blocked = proj.unsqueeze(1) # Swap Batch and channel dimensions
proj = (proj - self.sensor_img_means[:, None, None]
) / self.sensor_img_stds[:, None, None]
proj = proj * proj_mask.float()
# get name and sequence
path_norm = os.path.normpath(scan_file)
path_split = path_norm.split(os.sep)
path_seq = path_split[-3]
path_name = path_split[-1].replace(".bin", ".label")
# print("path_norm: ", path_norm)
# print("path_seq", path_seq)
# print("path_name", path_name)
# import time
# import cv2
# cv2.imwrite('/home/snowflake/Desktop/big8192-128.png', proj_blocked[0,0, :, :].numpy()*15)
# print('proj_blocked.shape')
# print(proj_blocked.shape)
# time.sleep(1000)
n, c, h, w = proj_blocked.size()
proj2 = proj.clone()
proj = proj.unsqueeze(0)
mask_image = proj_mask.unsqueeze(0).unsqueeze(0).float()
downsamplings = 4
representations = {}
representations['image'] = []
representations['points'] = []
windows_size = 3 # windows size
for i in range(downsamplings):
proj_chan_group_points = f.unfold(proj_blocked,
kernel_size=3,
stride=1,
padding=1)
projmask_chan_group_points = f.unfold(mask_image,
kernel_size=3,
stride=1,
padding=1)
# Get the mean point (taking apart non-valid points)
proj_chan_group_points_sum = torch.sum(proj_chan_group_points,
dim=1)
projmask_chan_group_points_sum = torch.sum(
projmask_chan_group_points, dim=1)
proj_chan_group_points_mean = proj_chan_group_points_sum / projmask_chan_group_points_sum
# tile it for being able to substract it to the other points
tiled_proj_chan_group_points_mean = proj_chan_group_points_mean.unsqueeze(
1).repeat(1, windows_size * windows_size, 1)
# remove nans due to empty blocks
is_nan = tiled_proj_chan_group_points_mean != tiled_proj_chan_group_points_mean
tiled_proj_chan_group_points_mean[is_nan] = 0.
# compute valid mask per point
tiled_projmask_chan_group_points = (
1 - projmask_chan_group_points.repeat(n, 1, 1)).byte()
# substract mean point to points
proj_chan_group_points_relative = proj_chan_group_points - tiled_proj_chan_group_points_mean
# set to zero points which where non valid at the beginning
proj_chan_group_points_relative[
tiled_projmask_chan_group_points] = 0.
# compute distance (radius) to mean point
# xyz_relative = proj_chan_group_points_relative[1:4,...]
# relative_distance = torch.norm(xyz_relative, dim=0).unsqueeze(0)
# NOW proj_chan_group_points_relative HAS Xr, Yr, Zr, Rr, Dr relative to the mean point
proj_norm_chan_group_points = f.unfold(proj.permute(1, 0, 2, 3),
kernel_size=3,
stride=1,
padding=1)
# NOW proj_norm_chan_group_points HAS X, Y, Z, R, D. Now we have to concat them both
proj_chan_group_points_combined = torch.cat(
[proj_norm_chan_group_points, proj_chan_group_points_relative],
dim=0)
# convert back to image for image-convolution-branch
proj_out = f.fold(proj_chan_group_points_combined,
proj_blocked.shape[-2:],
kernel_size=3,
stride=1,
padding=1)
proj_out = proj_out.squeeze(1)
proj = nn.functional.interpolate(proj,
size=(int(proj.shape[2] / 2),
int(proj.shape[3] / 2)),
mode='nearest')
proj_blocked = nn.functional.interpolate(
proj_blocked.permute(1, 0, 2, 3),
size=(int(proj_blocked.shape[2] / 2),
int(proj_blocked.shape[3] / 2)),
mode='nearest').permute(1, 0, 2, 3)
mask_image = nn.functional.interpolate(
mask_image,
size=(int(mask_image.shape[2] / 2),
int(mask_image.shape[3] / 2)),
mode='nearest')
representations['points'].append(proj_chan_group_points_combined)
representations['image'].append(proj_out)
# print('append' +str(i))
#
# print(proj_chan_group_points_combined.shape)
# print(proj_out.shape)
return proj2, proj_mask, proj_labels, unproj_labels, path_seq, path_name, proj_x, proj_y, proj_range, unproj_range, proj_xyz, unproj_xyz, proj_remission, unproj_remissions, unproj_n_points, representations
