for start, end in training_batch:
subdata = imgs_train[start:end].copy()
subcom = coms_train[start:end].copy()
subcube = cubes_train[start:end].copy()
subM = Ms_train[start:end].copy()
subgt3Dcrop = gt3Dcrops_train[start:end].copy()
resdata = np.ones_like(subdata)
resgt3D = np.ones_like(subgt3Dcrop)
hts = np.zeros(shape=(batch_size, 24, 24, 14))
jtIs = np.zeros(shape=(batch_size, 42))
for idx in range(batch_size):
dm = norm_dm(subdata[idx], subcom[idx], subcube[idx])
s = augmentCrop(dm,
subgt3Dcrop[idx],
di_train.joint3DToImg(subcom[idx]),
subcube[idx],
subM[idx], ['com', 'rot', 'none', 'sc'],
hd,
False,
rng=rng)
resdata[idx] = s[0]
resgt3D[idx] = s[2]
mode = s[7]
gt3D_ = resgt3D[idx] * subcube[idx][0] / 2. + subcom[idx]
jtI_ = transformPoints2D(di_train.joints3DToImg(gt3D_),
subM[idx])
jtI_ = np.reshape(jtI_, (1, 42))
ht_ = joints_heatmap_gen([1], jtI_, (24, 24), points=14)
hts[idx] = np.transpose(ht_, (0, 2, 3, 1)) / 255.
sess.run(lr_update)
for start, end in training_batch:
subdata = imgs_train[start:end].copy()
subcom = coms_train[start:end].copy()
subcube = cubes_train[start:end].copy()
subM = Ms_train[start:end].copy()
subgt3Dcrop = gt3Dcrops_train[start:end].copy()
resdata=np.ones_like(subdata)
resgt3D=np.ones_like(subgt3Dcrop)
hts=np.zeros(shape=(batch_size,24,24,14))
jtIs=np.zeros(shape=(batch_size,42))
for idx in range(batch_size):
dm = norm_dm(subdata[idx], subcom[idx], subcube[idx])
s = augmentCrop(dm, subgt3Dcrop[idx], di_train.joint3DToImg(subcom[idx]),
subcube[idx], subM[idx],['com', 'rot', 'none','sc'], hd, False, rng=rng)
resdata[idx] = s[0]
resgt3D[idx] = s[2]
mode=s[7]
gt3D_ = resgt3D[idx] * subcube[idx][0] / 2. + subcom[idx]
jtI_ = transformPoints2D(di_train.joints3DToImg(gt3D_), subM[idx])
jtI_=np.reshape(jtI_,(1,42))
ht_=joints_heatmap_gen([1],jtI_,(24,24),points=14)
hts[idx]=np.transpose(ht_,(0,2,3,1))/255.
feed_dict = {inputs: resdata.reshape(-1, 96, 96, 1),
label: resgt3D.reshape(-1, 42),
gt_ht:hts,
is_train:True,
class dataset_hand_NYU(data.Dataset):
def __init__(self, specs):
seed = specs['seed']
root = specs['root']
subset = specs['subset']
docom = specs['docom']
self.rng = np.random.RandomState(seed)
self.sampled_poses = None
self.pose_only = False
self.nmax = np.inf
self.augment = specs['augment']
self.num_sample_poses = specs['sample_poses']
self.joint_subset = specs['joint_subset']
self.aug_modes = ['none', 'com', 'rot']
print("create data")
self.flip_y = False
com_idx = 32
cube_size = 300
if 'MSRA' in self.joint_subset:
self.joint_subset = np.asarray([29, 23, 22, 20, 18, 17, 16, 14, 12, 11, 10,\
8, 6, 5, 4, 2, 0, 28, 27, 25, 24], dtype='int32')
com_idx = 17
elif 'ICVL' in self.joint_subset:
self.joint_subset = np.asarray([34, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10,\
8, 6, 4, 2, 0], dtype='int32')
self.flip_y = True
com_idx = 34
cube_size = 350
else:
self.joint_subset = np.arange(36)
self.di = NYUImporter(root,
refineNet=None,
allJoints=True,
com_idx=com_idx,
cacheDir='../../cache/')
if 'synth' in subset:
self.di.default_cubes[subset] = (cube_size, cube_size, cube_size)
print(self.di.default_cubes[subset])
self.Seq = self.di.loadSequence(subset,
rng=self.rng,
shuffle=True,
docom=docom)
#print(self.Seq.data[0].gt3Dcrop)
#self.di.showAnnotatedDepth(self.Seq.data[0])
#print('joint_subset', self.joint_subset)
# create training data
cube = np.asarray(self.Seq.config['cube'], 'float32')
com = np.asarray(self.Seq.data[0].com, 'float32')
img = np.asarray(self.Seq.data[0].dpt.copy(), 'float32')
img = normalize(img, com, cube)
self.hd = HandDetector(img,
abs(self.di.fx),
abs(self.di.fy),
importer=self.di)
self.num = len(self.Seq.data)
print(' data loaded with %d samples' % self.num)
def sample_poses(self):
train_cube = np.asarray([self.Seq.config['cube']] * self.num,
dtype='float32')
train_com = np.asarray([d.com for d in self.Seq.data], dtype='float32')
train_gt3D = np.asarray([d.gt3Dcrop for d in self.Seq.data],
dtype='float32')
self.sampled_poses = self.hd.sampleRandomPoses(self.di, self.rng, train_gt3D, train_com,\
train_cube, self.num_sample_poses, self.nmax, \
self.aug_modes)#.reshape((-1, train_gt3D.shape[1]*3))
self.num = self.sampled_poses.shape[0]
self.nmax = self.sampled_poses.shape[0]
print('%d sample poses created!' % self.num)
def __getitem__(self, i):
if self.pose_only and self.sampled_poses is not None:
pos = self.sampled_poses[i][self.joint_subset]
if self.flip_y:
pos[:, 1] *= -1
return pos.flatten()
cube = np.asarray(self.Seq.config['cube'], 'float32')
com = np.asarray(self.Seq.data[i].com, 'float32')
M = np.asarray(self.Seq.data[i].T, dtype='float32')
gt3D = np.asarray(self.Seq.data[i].gt3Dcrop, dtype='float32')
img = np.asarray(self.Seq.data[i].dpt.copy(), 'float32')
img = normalize(img, com, cube)
if not self.augment:
if self.joint_subset is not None:
gt3D = gt3D[self.joint_subset]
if self.flip_y:
gt3D[:, 1] *= -1
if self.pose_only:
return gt3D.flatten() / (cube[2] / 2.)
#print(img.shape, gt3D.flatten().shape, com.shape, M.shape, cube.shape)
return np.expand_dims(
img,
axis=0), gt3D.flatten() / (cube[2] / 2.), com, M, cube, cube
img, _, gt3D, cube, com2D, M, _ = augmentCrop(img, gt3D, \
self.di.joint3DToImg(com), cube, M, self.aug_modes, self.hd, rng=self.rng)
if self.joint_subset is not None:
gt3D = gt3D[self.joint_subset]
if self.flip_y:
gt3D[:, 1] *= -1
if self.pose_only:
return gt3D.flatten()
#print(imgD.shape, gt3Dn.flatten().shape, com.shape, M.shape, cube.shape)
return np.expand_dims(
img,
axis=0), gt3D.flatten(), self.di.jointImgTo3D(com2D), M, cube, cube
def set_nmax(self, frac):
self.nmax = int(self.num * frac)
print('self.nmax %d' % self.nmax)
def __len__(self):
return np.minimum(self.num, self.nmax)