def fn(elem):
dm, cfg, com = elem[0], elem[1], elem[2]
zz = tf.squeeze(dm, axis=-1)
min_depth = com[2] - D_RANGE * 0.5
max_depth = com[2] + D_RANGE * 0.5
zz = tf.where(tf.less(zz, -0.99),
tf.ones_like(zz) * max_depth,
tf.multiply(zz, D_RANGE) + min_depth)
xx, yy = tf.meshgrid(tf.range(h), tf.range(w))
xx = tf.to_float(xx)
yy = tf.to_float(yy)
w_ratio = cfg[4] / w
h_ratio = cfg[5] / h
new_cfg = CameraConfig(cfg[0] / w_ratio, cfg[1] / h_ratio,
cfg[2] / w_ratio, cfg[3] / h_ratio, w, h)
xx = tf.multiply(xx - new_cfg[2], tf.divide(zz, new_cfg[0]))
yy = tf.multiply(yy - new_cfg[3], tf.divide(zz, new_cfg[1]))
# renormalize the points as normalizing the pose
xx = tf.divide(xx - com[0], POSE_NORM_RATIO)
yy = tf.divide(yy - com[1], POSE_NORM_RATIO)
zz = tf.divide(zz - com[2], POSE_NORM_RATIO)
xyz = tf.stack([xx, yy, zz], axis=-1)
return [xyz, cfg, com]
def to_uvd_fn(elem):
xyz_pt, cfg = elem[0], elem[1]
return [
data.util.xyz2uvd_op(xyz_pt,
CameraConfig(*tf.unstack(cfg, axis=0))),
cfg
]
def fn(elems):
xyz_pose, cfg = elems[0], elems[1]
w_ratio = cfg[4] / out_w
h_ratio = cfg[5] / out_h
new_cfg = CameraConfig(cfg[0] / w_ratio, cfg[1] / h_ratio,
cfg[2] / w_ratio, cfg[3] / h_ratio, out_w,
out_h)
xx, yy = tf.meshgrid(tf.range(out_h), tf.range(out_w))
xx, yy = tf.cast(xx, tf.float32), tf.cast(yy, tf.float32)
xx = tf.tile(tf.expand_dims(xx, axis=-1), [1, 1, self._jnt_num])
yy = tf.tile(tf.expand_dims(yy, axis=-1), [1, 1, self._jnt_num])
uvd_pose = tf.reshape(data.util.xyz2uvd_op(xyz_pose, new_cfg),
(-1, 3))
[uu, vv, dd] = tf.unstack(uvd_pose, axis=-1)
uu = tf.reshape(uu, (1, 1, -1))
vv = tf.reshape(vv, (1, 1, -1))
hm = tf.maximum(
self.max_dist_2d -
tf.sqrt(tf.square(xx - uu) + tf.square(yy - vv)),
tf.zeros_like(xx)) / self.max_dist_2d
return [hm, cfg]
def fn(elems):
xyz_pt, com, cfg, hm, dm = elems[0], elems[1], elems[2], elems[
3], elems[4]
xx, yy, zz = tf.unstack(tf.reshape(xyz_pt, (-1, 3)), axis=-1)
xyz_pt = tf.reshape(xyz_pt, (-1, ))
xyz_pt = tf.multiply(xyz_pt,
data.preprocess.POSE_NORM_RATIO) + tf.tile(
com, [jnt_num * pnt_num])
xyz_pt = tf.reshape(xyz_pt, (-1, 3))
w_ratio = cfg[4] / out_w
h_ratio = cfg[5] / out_h
new_cfg = CameraConfig(cfg[0] / w_ratio, cfg[1] / h_ratio,
cfg[2] / w_ratio, cfg[3] / h_ratio, out_w,
out_h)
uvd_pt = xyz2uvd_op(xyz_pt, new_cfg)
uvd_pt = tf.reshape(uvd_pt, (-1, 3))
uu, vv, dd = tf.unstack(uvd_pt, axis=-1)
uu = tf.to_int32(uu + 0.5)
vv = tf.to_int32(vv + 0.5)
jj = tf.tile(tf.expand_dims(tf.range(jnt_num), axis=-1),
[1, pnt_num])
jj = tf.reshape(jj, (-1, ))
indices = tf.stack([vv, uu, jj], axis=-1)
weights = tf.gather_nd(hm, indices)
weights = tf.reshape(weights, (jnt_num, pnt_num, 1))
#we also clip the values of depth
dm = tf.squeeze(dm)
dm = tf.divide(
dm * data.preprocess.D_RANGE - data.preprocess.D_RANGE * 0.5,
data.preprocess.POSE_NORM_RATIO)
indices = tf.stack([vv, uu], axis=-1)
od = tf.gather_nd(dm, indices)
zz = tf.maximum(zz, od)
xyz_pt = tf.stack([xx, yy, zz], axis=-1)
xyz_pt = tf.reshape(xyz_pt, (jnt_num, pnt_num, 3))
return [weights, xyz_pt, cfg, hm, dm]