def normalize(xyzs, rgbs, covars, lbls, offset_z, bsize=3.0):
bn = len(xyzs)
block_mins = []
block_nidxs = []
block_nlens = []
block_nbegs = []
# t=0
for bid in xrange(bn):
min_xyz = np.min(xyzs[bid], axis=0, keepdims=True)
min_xyz[:, 2] = offset_z
min_xyz[:, :2] += bsize / 2.0
xyzs[bid] -= min_xyz
block_mins.append(min_xyz)
rgbs[bid][:, :3] -= 128
rgbs[bid][:, :3] /= 128
# sub mean div std
if rgbs[bid].shape[1] > 3:
rgbs[bid][:, 3] -= -1164.05
rgbs[bid][:, 3] /= 600.0
nidxs = libPointUtil.findNeighborRadiusCPU(xyzs[bid], radius)
nlens = np.asarray([len(idxs) for idxs in nidxs], dtype=np.int32)
nbegs = compute_nidxs_bgs(nlens)
nidxs = np.concatenate(nidxs, axis=0)
block_nidxs.append(nidxs)
block_nlens.append(nlens)
block_nbegs.append(nbegs)
return xyzs, rgbs, covars, lbls, block_mins, block_nidxs, block_nlens, block_nbegs
def normalize_model_hierarchy(xyzs,
use_rotate=False,
nr1=0.05,
nr2=0.1,
nr3=0.3,
vc1=0.1,
vc2=0.3):
#xyzs [b,1024,3]
n = xyzs.shape[0]
cxyzs,dxyzs,vlens,vlens_bgs,vcidxs,\
cidxs,nidxs,nidxs_bgs,nidxs_lens=[],[],[],[],[],[],[],[],[]
covars = []
for i in xrange(n):
# bg=time.time()
if use_rotate:
ang = np.random.uniform(0, np.pi * 2)
xyzs[i] = rotate(xyzs[i], ang)
nidxs1 = libPointUtil.findNeighborRadiusCPU(xyzs[i], nr1, 15)
nidxs_lens1 = np.ascontiguousarray([len(idxs) for idxs in nidxs1],
np.int32)
nidxs_bgs1 = compute_nidxs_bgs(nidxs_lens1)
nidxs1 = np.ascontiguousarray(np.concatenate(nidxs1, axis=0),
dtype=np.int32)
covar = libPointUtil.computeCovarsGPU(xyzs[i], nidxs1, nidxs_lens1,
nidxs_bgs1)
cxyz1, dxyz1, vlens1, cxyz2, dxyz2, vlens2, cxyz3, feats_list=\
build_hierarchy(xyzs[i],[covar],vc1,vc2)
covar = feats_list[0]
covars.append(covar)
# rescale to unit cube
dxyz1 /= vc1
dxyz2 /= vc2
# bg=time.time()
vlens_bgs1 = compute_nidxs_bgs(vlens1)
vlens_bgs2 = compute_nidxs_bgs(vlens2)
vcidxs1 = compute_cidxs(vlens1)
vcidxs2 = compute_cidxs(vlens2)
nidxs1, nidxs_lens1, nidxs_bgs1, cidxs1 = point2ndixs(cxyz1, nr1)
nidxs2, nidxs_lens2, nidxs_bgs2, cidxs2 = point2ndixs(cxyz2, nr2)
nidxs3, nidxs_lens3, nidxs_bgs3, cidxs3 = point2ndixs(cxyz3, nr3)
_append([
cxyzs, dxyzs, vlens, vlens_bgs, vcidxs, cidxs, nidxs, nidxs_bgs,
nidxs_lens
], [[cxyz1, cxyz2, cxyz3], [dxyz1, dxyz2], [vlens1, vlens2],
[vlens_bgs1, vlens_bgs2], [vcidxs1, vcidxs2],
[cidxs1, cidxs2, cidxs3], [nidxs1, nidxs2, nidxs3],
[nidxs_bgs1, nidxs_bgs2, nidxs_bgs3],
[nidxs_lens1, nidxs_lens2, nidxs_lens3]])
# print '{} cost {} s'.format(i,time.time()-bg)
return cxyzs, dxyzs, covars, vlens, vlens_bgs, vcidxs, cidxs, nidxs, nidxs_bgs, nidxs_lens
def test_nn():
xyzs = np.random.uniform(-1, 1, [1024, 3])
xyzs = np.asarray(xyzs, np.float32)
covar_nidxs = libPointUtil.findNeighborRadiusCPU(xyzs, radius)
for k in xrange(1024):
for pi, pt in enumerate(xyzs):
if np.sum((pt - xyzs[k])**2, axis=0) < radius**2:
assert pi in covar_nidxs[k]
for ni in covar_nidxs[k]:
assert np.sum((xyzs[ni] - xyzs[k])**2, axis=0) < radius**2
def compute_covar(xyzs, radius):
covar_nidxs = libPointUtil.findNeighborRadiusCPU(xyzs, radius)
covar_nidxs_lens = np.ascontiguousarray(
[len(idxs) for idxs in covar_nidxs], np.int32)
covar_nidxs_bgs = compute_nidxs_bgs(covar_nidxs_lens)
covar_nidxs = np.ascontiguousarray(np.concatenate(covar_nidxs, axis=0),
dtype=np.int32)
covars = libPointUtil.computeCovarsGPU(xyzs, covar_nidxs, covar_nidxs_lens,
covar_nidxs_bgs)
return covars
def point2ndixs(pts, radius):
leafsize = 10 if pts.shape[0] > 100 else 1
if pts.shape[0] > 8000:
nidxs = libPointUtil.findNeighborRadiusGPU(pts, radius, leafsize)
else:
nidxs = libPointUtil.findNeighborRadiusCPU(pts, radius, leafsize)
nidxs_lens = [len(nidx) for nidx in nidxs]
nidxs_lens = np.asarray(nidxs_lens, dtype=np.int32)
nidxs = np.concatenate(nidxs, axis=0)
nidxs_bgs = compute_nidxs_bgs(nidxs_lens)
cindxs = compute_cidxs(nidxs_lens)
return nidxs, nidxs_lens, nidxs_bgs, cindxs
def sample_block(points,
labels,
ds_stride,
block_size,
block_stride,
min_pn,
use_rescale=False,
use_flip=False,
use_rotate=False,
covar_ds_stride=0.03,
covar_nn_size=0.1,
gpu_gather=False):
xyzs = np.ascontiguousarray(points[:, :3])
rgbs = np.ascontiguousarray(points[:, 3:])
min_xyzs = np.min(xyzs, axis=0, keepdims=True)
max_xyzs = np.max(xyzs, axis=0, keepdims=True)
covar_ds_idxs = libPointUtil.gridDownsampleGPU(xyzs, covar_ds_stride,
False)
covar_ds_xyzs = np.ascontiguousarray(xyzs[covar_ds_idxs, :])
# flip
if use_flip:
if random.random() < 0.5:
covar_ds_xyzs = swap_xy(covar_ds_xyzs)
min_xyzs = swap_xy(min_xyzs)
max_xyzs = swap_xy(max_xyzs)
if random.random() < 0.5:
covar_ds_xyzs = flip(covar_ds_xyzs, axis=0)
min_xyzs[:, 0], max_xyzs[:, 0] = -max_xyzs[:, 0], -min_xyzs[:, 0]
if random.random() < 0.5:
covar_ds_xyzs = flip(covar_ds_xyzs, axis=1)
min_xyzs[:, 1], max_xyzs[:, 1] = -max_xyzs[:, 1], -min_xyzs[:, 1]
# rescale
if use_rescale:
rescale = np.random.uniform(0.9, 1.1, [1, 3])
covar_ds_xyzs[:, :3] *= rescale
min_xyzs *= rescale
max_xyzs *= rescale
# rotate
if use_rotate:
if random.random() > 0.3:
angle = random.random() * np.pi / 2.0
covar_ds_xyzs = rotate(covar_ds_xyzs, angle)
ds_idxs = libPointUtil.gridDownsampleGPU(covar_ds_xyzs, ds_stride, False)
covar_nidxs = libPointUtil.findNeighborRadiusCPU(covar_ds_xyzs, ds_idxs,
covar_nn_size)
covar_nidxs_lens = np.ascontiguousarray(
[len(idxs) for idxs in covar_nidxs], np.int32)
covar_nidxs_bgs = compute_nidxs_bgs(covar_nidxs_lens)
covar_nidxs = np.ascontiguousarray(np.concatenate(covar_nidxs, axis=0),
dtype=np.int32)
covars = libPointUtil.computeCovarsGPU(covar_ds_xyzs, covar_nidxs,
covar_nidxs_lens, covar_nidxs_bgs)
xyzs = covar_ds_xyzs[ds_idxs, :]
rgbs = rgbs[covar_ds_idxs, :][ds_idxs, :]
lbls = labels[covar_ds_idxs][ds_idxs]
if gpu_gather:
xyzs -= min_xyzs
idxs = uniform_sample_block_gpu(xyzs,
block_size,
block_stride,
min_pn=min_pn)
xyzs += min_xyzs
else:
xyzs -= min_xyzs
idxs = uniform_sample_block(xyzs,
block_size,
block_stride,
normalized=True,
min_pn=min_pn)
xyzs += min_xyzs
xyzs, rgbs, covars, lbls = fetch_subset([xyzs, rgbs, covars, lbls], idxs)
return xyzs, rgbs, covars, lbls
def sample_context_block(tfs, points, labels, global_points, ds_stride, block_size, block_stride, min_pn,
use_rescale=False, use_flip=False, covar_ds_stride=0.03, covar_nn_size=0.1,
context_len=50.0):
xyzs=np.ascontiguousarray(points[:,:3])
rgbs=np.ascontiguousarray(points[:,3:])
min_xyz=np.min(xyzs,axis=0,keepdims=True)
max_xyz=np.max(xyzs,axis=0,keepdims=True)
covar_ds_idxs=libPointUtil.gridDownsampleGPU(xyzs, covar_ds_stride, False)
covar_ds_xyzs=np.ascontiguousarray(xyzs[covar_ds_idxs,:])
# flip
if use_flip:
if random.random()<0.5:
covar_ds_xyzs=swap_xy(covar_ds_xyzs)
global_points=swap_xy(global_points)
min_xyz=swap_xy(min_xyz)
max_xyz=swap_xy(max_xyz)
if random.random()<0.5:
covar_ds_xyzs=flip(covar_ds_xyzs,axis=0)
global_points=flip(global_points,axis=0)
min_xyz[:,0],max_xyz[:,0]=-max_xyz[:,0],-min_xyz[:,0]
if random.random()<0.5:
covar_ds_xyzs=flip(covar_ds_xyzs,axis=1)
global_points=flip(global_points,axis=1)
min_xyz[:,1],max_xyz[:,1]=-max_xyz[:,1],-min_xyz[:,1]
# rescale
if use_rescale:
rescale=np.random.uniform(0.9,1.1,[1,3])
covar_ds_xyzs[:,:3]*=rescale
global_points[:,:3]*=rescale
min_xyz*=rescale
max_xyz*=rescale
ds_idxs=libPointUtil.gridDownsampleGPU(covar_ds_xyzs,ds_stride,False)
# compute covar
covar_nidxs=libPointUtil.findNeighborRadiusCPU(covar_ds_xyzs,ds_idxs,covar_nn_size)
covar_nidxs_lens=np.ascontiguousarray([len(idxs) for idxs in covar_nidxs],np.int32)
covar_nidxs_bgs=compute_nidxs_bgs(covar_nidxs_lens)
covar_nidxs=np.ascontiguousarray(np.concatenate(covar_nidxs,axis=0),dtype=np.int32)
covars=libPointUtil.computeCovarsGPU(covar_ds_xyzs,covar_nidxs,covar_nidxs_lens,covar_nidxs_bgs)
xyzs=covar_ds_xyzs[ds_idxs,:]
rgbs=rgbs[covar_ds_idxs,:][ds_idxs,:]
lbls=labels[covar_ds_idxs][ds_idxs]
xyzs-=min_xyz
idxs=uniform_sample_block(xyzs,block_size,block_stride,min_pn=min_pn)
xyzs+=min_xyz
xyzs, rgbs, covars, lbls=fetch_subset([xyzs,rgbs,covars,lbls],idxs)
context_xyzs=compute_context_xyzs(global_points,xyzs,context_len=context_len)
for ci,ctx_xyz in enumerate(context_xyzs):
if ctx_xyz.shape[0]==0:
print '!!!! error {}'.format(tfs)
raise RuntimeError
context_idxs=compute_context_idxs(context_xyzs,xyzs)
return xyzs, rgbs, covars, lbls, context_xyzs, context_idxs