def untransform(self, points):
points = ensure_homogeneous(points, d=3)
if len(self.shuffle_indices) > 0:
inv_shuffle_indices = list(range(len(self.shuffle_indices)))
for i, j in enumerate(self.shuffle_indices):
inv_shuffle_indices[j] = i
points = points[..., inv_shuffle_indices, :]
return points @ cast_array(np.linalg.inv(self.matrix).T, points)
def test_normalise_points(self, points, camera, z_ref, skeleton):
denorm_pose = ensure_homogeneous(points.copy(), d=3)
denorm_pose[:, :2] -= denorm_pose[skeleton.root_joint_id, :2]
norm_pose = normalise_points(denorm_pose, z_ref, camera, 2048, 2048)
assert_allclose(norm_pose[skeleton.root_joint_id],
np.asarray([0.0, 0.0, 0.0, 1.0]))
actual = norm_pose[1]
expected = torch.as_tensor([0.0215, -0.1514, -0.0127, 1.0000])
assert_allclose(actual, expected, rtol=0, atol=1e-4)
def test_infer_z_depth(self, points, camera, z_ref, skeleton):
denorm_pose = ensure_homogeneous(points.copy(), d=3)
denorm_pose[:, :2] -= denorm_pose[skeleton.root_joint_id, :2]
norm_pose = normalise_points(denorm_pose, z_ref, camera, 2048, 2048)
right_wrist = skeleton.joint_index('right_wrist')
right_elbow = skeleton.joint_index('right_elbow')
target_forearm_length = np.linalg.norm(denorm_pose[right_wrist] -
denorm_pose[right_elbow])
def eval_scale(skel):
forearm_length = np.linalg.norm(skel[right_wrist] -
skel[right_elbow])
return forearm_length / target_forearm_length
inferred = infer_z_depth(norm_pose, eval_scale, camera, 2048, 2048)
assert inferred == pytest.approx(z_ref, rel=0, abs=1e-2)
def project_cartesian(self, coords):
coords = ensure_homogeneous(coords, d=3)
return to_cartesian(self.project(coords))
def test_denormalise_points(self, points, camera, z_ref, skeleton):
denorm_pose = ensure_homogeneous(points.copy(), d=3)
denorm_pose[:, :2] -= denorm_pose[skeleton.root_joint_id, :2]
norm_pose = normalise_points(denorm_pose, z_ref, camera, 2048, 2048)
recons_pose = denormalise_points(norm_pose, z_ref, camera, 2048, 2048)
assert_allclose(recons_pose, denorm_pose, rtol=0, atol=1e-4)
def transform(self, points):
points = ensure_homogeneous(points, d=3)
if len(self.shuffle_indices) > 0:
points = points[..., self.shuffle_indices, :]
return points @ cast_array(self.matrix.T, points)