# transform point clouds back to their reference frame using ground truth
# (this is important because the network must learn when point clouds are in their original reference frame)
pcd1_corr.transform(np.linalg.inv(T1))
pcd2_corr.transform(np.linalg.inv(T2))
# extract patches and compute LRFs
patches1_batch = np.empty((batch_size, 3, patch_size))
patches2_batch = np.empty((batch_size, 3, patch_size))
lrfs1_batch = np.empty((batch_size, 4, 4))
lrfs2_batch = np.empty((batch_size, 4, 4))
frag1_lrf = lrf(pcd=pcd1,
pcd_tree=o3d.geometry.KDTreeFlann(pcd1),
patch_size=patch_size,
lrf_kernel=lrf_kernel,
viz=False)
frag2_lrf = lrf(pcd=pcd2,
pcd_tree=o3d.geometry.KDTreeFlann(pcd2),
patch_size=patch_size,
lrf_kernel=lrf_kernel,
viz=False)
# GET PATCHES USING THE FPS POINTS
for i in range(len(fps_pcd1_idx)):
pt1 = np.asarray(pcd1_corr.points)[fps_pcd1_idx.cpu()[i]]
pt2 = np.asarray(pcd2_corr.points)[fps_pcd2_idx.cpu()[i]]
np.asarray(pcd1.points).shape[0]))
inds2 = np.random.choice(np.asarray(pcd2.points).shape[0],
pts_to_sample,
replace=False)
print('randomly sampled {} out of {} from pcd2 to compute DIPs'.format(
pts_to_sample,
np.asarray(pcd2.points).shape[0]))
pcd1_pts = np.asarray(pcd1.points)[inds1]
pcd2_pts = np.asarray(pcd2.points)[inds2]
# compute DIPs
frag1_lrf = lrf(
pcd=pcd1,
pcd_tree=o3d.geometry.KDTreeFlann(pcd1),
patch_size=patch_size,
lrf_kernel=lrf_kernel,
viz=False
) # set viz=True to visualise patch kernels and their respective LRF
frag2_lrf = lrf(
pcd=pcd2,
pcd_tree=o3d.geometry.KDTreeFlann(pcd2),
patch_size=patch_size,
lrf_kernel=lrf_kernel,
viz=False
) # set viz=True to visualise patch kernels and their respective LRF
patches1 = np.empty((pcd1_pts.shape[0], 3, patch_size))
patches2 = np.empty((pcd2_pts.shape[0], 3, patch_size))