def loader(filename, test_angle=None):
P = np.asarray(open3d.read_point_cloud(filename).points)
cls = classmap['_'.join(os.path.basename(filename).split('_')[:-1])]
#transform into ball of diameter 32 (obj scale in modelnet has no meaning, original meshes have random sizes)
# (in the paper we used a unit ball and ./32 grid sizes, this is equivalent in effect)
diameter = np.max(np.max(P, axis=0) - np.min(P, axis=0))
M = transforms3d.zooms.zfdir2mat(32 / diameter)
# training data augmentation
if training:
if args.pc_augm_input_dropout > 0: # removing points here changes graph structure (unlike zeroing features)
P, _ = pcu.dropout(P, None, args.pc_augm_input_dropout)
if args.pc_augm_scale > 1:
s = random.uniform(1 / args.pc_augm_scale, args.pc_augm_scale)
M = np.dot(transforms3d.zooms.zfdir2mat(s), M)
if args.pc_augm_rot:
angle = random.uniform(0, 2 * math.pi)
M = np.dot(transforms3d.axangles.axangle2mat([0, 0, 1], angle),
M) # z=upright assumption
if args.pc_augm_mirror_prob > 0: # mirroring x&y, not z
if random.random() < args.pc_augm_mirror_prob / 2:
M = np.dot(transforms3d.zooms.zfdir2mat(-1, [1, 0, 0]), M)
if random.random() < args.pc_augm_mirror_prob / 2:
M = np.dot(transforms3d.zooms.zfdir2mat(-1, [0, 1, 0]), M)
else:
if test_angle:
M = np.dot(
transforms3d.axangles.axangle2mat([0, 0, 1], test_angle),
M) # z=upright assumption
P = np.dot(P, M.T)
# coarsen to initial resolution (btw, axis-aligned quantization of rigidly transformed cloud adds jittering noise)
P -= np.min(
P, axis=0
) #move to positive octant (voxelgrid has fixed boundaries at axes planes)
cloud = pcu.create_cloud(P)
cloud = open3d.voxel_down_sample(cloud, voxel_size=pyramid_conf[0][0])
F = np.ones((len(cloud.points), 1),
dtype=np.float32) # no point features in modelnet
graphs, poolmaps = pcu.create_graph_pyramid(args, cloud, pyramid_conf)
return F, cls, graphs, poolmaps
def loader(filename):
data = np.fromfile(filename, binType)
cls = classmap[os.path.basename(filename).split('.')[0]]
P = np.vstack([data['x'], data['y'], data['z']]).T # metric units
F = data['intensity'].reshape(-1, 1)
# training data augmentation
if training:
if args.pc_augm_input_dropout > 0: # removing points here changes graph structure (unlike zeroing features)
P, F = pcu.dropout(P, F, args.pc_augm_input_dropout)
M = np.eye(3)
if args.pc_augm_scale > 1:
s = random.uniform(1 / args.pc_augm_scale, args.pc_augm_scale)
M = np.dot(transforms3d.zooms.zfdir2mat(s), M)
if args.pc_augm_rot:
angle = random.uniform(0, 2 * math.pi)
M = np.dot(transforms3d.axangles.axangle2mat([0, 0, 1], angle),
M) # z=upright assumption
if args.pc_augm_mirror_prob > 0: # mirroring x&y, not z
if random.random() < args.pc_augm_mirror_prob / 2:
M = np.dot(transforms3d.zooms.zfdir2mat(-1, [1, 0, 0]), M)
if random.random() < args.pc_augm_mirror_prob / 2:
M = np.dot(transforms3d.zooms.zfdir2mat(-1, [0, 1, 0]), M)
P = np.dot(P, M.T)
# coarsen to initial resolution (btw, axis-aligned quantization of rigidly transformed cloud adds jittering noise)
P -= np.min(
P, axis=0
) #move to positive octant (voxelgrid has fixed boundaries at axes planes)
PF = np.hstack([P, F]).astype(np.float32)
PF_filtered = pcu.voxelgrid(
pcl.PointCloud_PointXYZI(PF), pyramid_conf[0][0]
).to_array(
) # aggregates intensities too (note pcl wrapper bug: only int intensities accepted)
F = PF_filtered[:,
3] / 255 - 0.5 # laser return intensities in [-0.5,0.5]
cloud = pcl.PointCloud(
PF_filtered[:, 0:3]
) # (pcl wrapper bug: XYZI cannot query kd-tree by radius)
graphs, poolmaps = pcu.create_graph_pyramid(args, cloud, pyramid_conf)
return F, cls, graphs, poolmaps
def loader(filename):
data = np.fromfile(filename, binType)
cls = classmap[os.path.basename(filename).split('.')[0]]
P = np.vstack([data['x'], data['y'], data['z']]).T # metric units
F = data['intensity'].reshape(-1, 1)
# training data augmentation
if training:
if args.pc_augm_input_dropout > 0: # removing points here changes graph structure (unlike zeroing features)
P, F = pcu.dropout(P, F, args.pc_augm_input_dropout)
M = np.eye(3)
if args.pc_augm_scale > 1:
s = random.uniform(1 / args.pc_augm_scale, args.pc_augm_scale)
M = np.dot(transforms3d.zooms.zfdir2mat(s), M)
if args.pc_augm_rot:
angle = random.uniform(0, 2 * math.pi)
M = np.dot(transforms3d.axangles.axangle2mat([0, 0, 1], angle),
M) # z=upright assumption
if args.pc_augm_mirror_prob > 0: # mirroring x&y, not z
if random.random() < args.pc_augm_mirror_prob / 2:
M = np.dot(transforms3d.zooms.zfdir2mat(-1, [1, 0, 0]), M)
if random.random() < args.pc_augm_mirror_prob / 2:
M = np.dot(transforms3d.zooms.zfdir2mat(-1, [0, 1, 0]), M)
P = np.dot(P, M.T)
# coarsen to initial resolution (btw, axis-aligned quantization of rigidly transformed cloud adds jittering noise)
P -= np.min(
P, axis=0
) #move to positive octant (voxelgrid has fixed boundaries at axes planes)
cloud = pcu.create_cloud(P, intensity=F)
cloud = open3d.voxel_down_sample(
cloud, voxel_size=pyramid_conf[0][0]) # aggregates intensities too
F = np.asarray(
cloud.colors
)[:, 0] / 255 - 0.5 # laser return intensities in [-0.5,0.5]
graphs, poolmaps = pcu.create_graph_pyramid(args, cloud, pyramid_conf)
return F.astype(np.float32), cls, graphs, poolmaps