class PixelMaskRenderer():
"""Render-ready scene composed of a single object and a single moving camera.
# Arguments
path_OBJ: String containing the path to an OBJ file.
viewport_size: List, specifying [H, W] of rendered image.
y_fov: Float indicating the vertical field of view in radians.
distance: List of floats indicating [max_distance, min_distance]
light: List of floats indicating [max_light, min_light]
top_only: Boolean. If True images are only take from the top.
roll: Float, to sample [-roll, roll] rolls of the Z OpenGL camera axis.
shift: Float, to sample [-shift, shift] to move in X, Y OpenGL axes.
"""
def __init__(self, path_OBJ, viewport_size=(128, 128), y_fov=3.14159 / 4.0,
distance=[0.3, 0.5], light=[0.5, 30], top_only=False,
roll=None, shift=None):
self.distance, self.roll, self.shift = distance, roll, shift
self.light_intensity, self.top_only = light, top_only
self._build_scene(path_OBJ, viewport_size, light, y_fov)
self.renderer = OffscreenRenderer(viewport_size[0], viewport_size[1])
self.flags_RGBA = RenderFlags.RGBA
self.flags_FLAT = RenderFlags.RGBA | RenderFlags.FLAT
self.epsilon = 0.01
def _build_scene(self, path, size, light, y_fov):
self.scene = Scene(bg_color=[0, 0, 0, 0])
self.light = self.scene.add(
DirectionalLight([1.0, 1.0, 1.0], np.mean(light)))
self.camera = self.scene.add(
PerspectiveCamera(y_fov, aspectRatio=np.divide(*size)))
self.pixel_mesh = self.scene.add(color_object(path))
self.mesh = self.scene.add(
Mesh.from_trimesh(trimesh.load(path), smooth=True))
self.world_origin = self.mesh.mesh.centroid
def _sample_parameters(self):
distance = sample_uniformly(self.distance)
camera_origin = sample_point_in_sphere(distance, self.top_only)
camera_origin = random_perturbation(camera_origin, self.epsilon)
light_intensity = sample_uniformly(self.light_intensity)
return camera_origin, light_intensity
def render(self):
camera_origin, intensity = self._sample_parameters()
camera_to_world, world_to_camera = compute_modelview_matrices(
camera_origin, self.world_origin, self.roll, self.shift)
self.light.light.intensity = intensity
self.scene.set_pose(self.camera, camera_to_world)
self.scene.set_pose(self.light, camera_to_world)
self.pixel_mesh.mesh.is_visible = False
image, depth = self.renderer.render(self.scene, self.flags_RGBA)
self.pixel_mesh.mesh.is_visible = True
image, alpha = split_alpha_channel(image)
self.mesh.mesh.is_visible = False
RGB_mask, _ = self.renderer.render(self.scene, self.flags_FLAT)
self.mesh.mesh.is_visible = True
return image, alpha, RGB_mask
class DictionaryView():
"""Render-ready scene composed of a single object and a single moving camera.
# Arguments
filepath: String containing the path to an OBJ file.
viewport_size: List, specifying [H, W] of rendered image.
y_fov: Float indicating the vertical field of view in radians.
distance: List of floats indicating [max_distance, min_distance]
top_only: Boolean. If True images are only take from the top.
light: List of floats indicating [max_light, min_light]
"""
def __init__(self,
filepath,
viewport_size=(128, 128),
y_fov=3.14159 / 4.,
distance=0.30,
top_only=False,
light=5.0,
theta_steps=10,
phi_steps=10):
self.scene = Scene(bg_color=[0, 0, 0])
self.camera = self.scene.add(
PerspectiveCamera(y_fov, aspectRatio=np.divide(*viewport_size)))
self.mesh = self.scene.add(
Mesh.from_trimesh(trimesh.load(filepath), smooth=True))
self.world_origin = self.mesh.mesh.centroid
self.light = self.scene.add(DirectionalLight([1.0, 1.0, 1.0], light))
self.distance = distance
# 0.1 values are to avoid gimbal lock
theta_max = np.pi / 2.0 if top_only else np.pi
self.thetas = np.linspace(0.1, theta_max - 0.1, theta_steps)
self.phis = np.linspace(0.1, 2 * np.pi - 0.1, phi_steps)
self.renderer = OffscreenRenderer(*viewport_size)
self.RGBA = RenderFlags.RGBA
def render(self):
dictionary_data = []
for theta_arg, theta in enumerate(self.thetas):
for phi_arg, phi in enumerate(self.phis):
x = self.distance * np.sin(theta) * np.cos(phi)
y = self.distance * np.sin(theta) * np.sin(phi)
z = self.distance * np.cos(theta)
matrices = compute_modelview_matrices(np.array([x, z, y]),
self.world_origin)
camera_to_world, world_to_camera = matrices
self.scene.set_pose(self.camera, camera_to_world)
self.scene.set_pose(self.light, camera_to_world)
camera_to_world = camera_to_world.flatten()
world_to_camera = world_to_camera.flatten()
image, depth = self.renderer.render(self.scene,
flags=self.RGBA)
image, alpha = split_alpha_channel(image)
matrices = np.vstack([world_to_camera, camera_to_world])
sample = {
'image': image,
'alpha': alpha,
'depth': depth,
'matrices': matrices
}
dictionary_data.append(sample)
return dictionary_data
def test_scenes():
# Basics
s = Scene()
assert np.allclose(s.bg_color, np.ones(4))
assert np.allclose(s.ambient_light, np.zeros(3))
assert len(s.nodes) == 0
assert s.name is None
s.name = 'asdf'
s.bg_color = None
s.ambient_light = None
assert np.allclose(s.bg_color, np.ones(4))
assert np.allclose(s.ambient_light, np.zeros(3))
assert s.nodes == set()
assert s.cameras == set()
assert s.lights == set()
assert s.point_lights == set()
assert s.spot_lights == set()
assert s.directional_lights == set()
assert s.meshes == set()
assert s.camera_nodes == set()
assert s.light_nodes == set()
assert s.point_light_nodes == set()
assert s.spot_light_nodes == set()
assert s.directional_light_nodes == set()
assert s.mesh_nodes == set()
assert s.main_camera_node is None
assert np.all(s.bounds == 0)
assert np.all(s.centroid == 0)
assert np.all(s.extents == 0)
assert np.all(s.scale == 0)
# From trimesh scene
tms = trimesh.load('tests/data/WaterBottle.glb')
s = Scene.from_trimesh_scene(tms)
assert len(s.meshes) == 1
assert len(s.mesh_nodes) == 1
# Test bg color formatting
s = Scene(bg_color=[0, 1.0, 0])
assert np.allclose(s.bg_color, np.array([0.0, 1.0, 0.0, 1.0]))
# Test constructor for nodes
n1 = Node()
n2 = Node()
n3 = Node()
nodes = [n1, n2, n3]
s = Scene(nodes=nodes)
n1.children.append(n2)
s = Scene(nodes=nodes)
n3.children.append(n2)
with pytest.raises(ValueError):
s = Scene(nodes=nodes)
n3.children = []
n2.children.append(n3)
n3.children.append(n2)
with pytest.raises(ValueError):
s = Scene(nodes=nodes)
# Test node accessors
n1 = Node()
n2 = Node()
n3 = Node()
nodes = [n1, n2]
s = Scene(nodes=nodes)
assert s.has_node(n1)
assert s.has_node(n2)
assert not s.has_node(n3)
# Test node poses
for n in nodes:
assert np.allclose(s.get_pose(n), np.eye(4))
with pytest.raises(ValueError):
s.get_pose(n3)
with pytest.raises(ValueError):
s.set_pose(n3, np.eye(4))
tf = np.eye(4)
tf[:3, 3] = np.ones(3)
s.set_pose(n1, tf)
assert np.allclose(s.get_pose(n1), tf)
assert np.allclose(s.get_pose(n2), np.eye(4))
nodes = [n1, n2, n3]
tf2 = np.eye(4)
tf2[:3, :3] = np.diag([-1, -1, 1])
n1.children.append(n2)
n1.matrix = tf
n2.matrix = tf2
s = Scene(nodes=nodes)
assert np.allclose(s.get_pose(n1), tf)
assert np.allclose(s.get_pose(n2), tf.dot(tf2))
assert np.allclose(s.get_pose(n3), np.eye(4))
n1 = Node()
n2 = Node()
n3 = Node()
n1.children.append(n2)
s = Scene()
s.add_node(n1)
with pytest.raises(ValueError):
s.add_node(n2)
s.set_pose(n1, tf)
assert np.allclose(s.get_pose(n1), tf)
assert np.allclose(s.get_pose(n2), tf)
s.set_pose(n2, tf2)
assert np.allclose(s.get_pose(n2), tf.dot(tf2))
# Test node removal
n1 = Node()
n2 = Node()
n3 = Node()
n1.children.append(n2)
n2.children.append(n3)
s = Scene(nodes=[n1, n2, n3])
s.remove_node(n2)
assert len(s.nodes) == 1
assert n1 in s.nodes
assert len(n1.children) == 0
assert len(n2.children) == 1
s.add_node(n2, parent_node=n1)
assert len(n1.children) == 1
n1.matrix = tf
n3.matrix = tf2
assert np.allclose(s.get_pose(n3), tf.dot(tf2))
# Now test ADD function
s = Scene()
m = Mesh([], name='m')
cp = PerspectiveCamera(yfov=2.0)
co = OrthographicCamera(xmag=1.0, ymag=1.0)
dl = DirectionalLight()
pl = PointLight()
sl = SpotLight()
n1 = s.add(m, name='mn')
assert n1.mesh == m
assert len(s.nodes) == 1
assert len(s.mesh_nodes) == 1
assert n1 in s.mesh_nodes
assert len(s.meshes) == 1
assert m in s.meshes
assert len(s.get_nodes(node=n2)) == 0
n2 = s.add(m, pose=tf)
assert len(s.nodes) == len(s.mesh_nodes) == 2
assert len(s.meshes) == 1
assert len(s.get_nodes(node=n1)) == 1
assert len(s.get_nodes(node=n1, name='mn')) == 1
assert len(s.get_nodes(name='mn')) == 1
assert len(s.get_nodes(obj=m)) == 2
assert len(s.get_nodes(obj=m, obj_name='m')) == 2
assert len(s.get_nodes(obj=co)) == 0
nsl = s.add(sl, name='sln')
npl = s.add(pl, parent_name='sln')
assert nsl.children[0] == npl
ndl = s.add(dl, parent_node=npl)
assert npl.children[0] == ndl
nco = s.add(co)
ncp = s.add(cp)
assert len(s.light_nodes) == len(s.lights) == 3
assert len(s.point_light_nodes) == len(s.point_lights) == 1
assert npl in s.point_light_nodes
assert len(s.spot_light_nodes) == len(s.spot_lights) == 1
assert nsl in s.spot_light_nodes
assert len(s.directional_light_nodes) == len(s.directional_lights) == 1
assert ndl in s.directional_light_nodes
assert len(s.cameras) == len(s.camera_nodes) == 2
assert s.main_camera_node == nco
s.main_camera_node = ncp
s.remove_node(ncp)
assert len(s.cameras) == len(s.camera_nodes) == 1
assert s.main_camera_node == nco
s.remove_node(n2)
assert len(s.meshes) == 1
s.remove_node(n1)
assert len(s.meshes) == 0
s.remove_node(nsl)
assert len(s.lights) == 0
s.remove_node(nco)
assert s.main_camera_node is None
s.add_node(n1)
s.clear()
assert len(s.nodes) == 0
# Trigger final errors
with pytest.raises(ValueError):
s.main_camera_node = None
with pytest.raises(ValueError):
s.main_camera_node = ncp
with pytest.raises(ValueError):
s.add(m, parent_node=n1)
with pytest.raises(ValueError):
s.add(m, name='asdf')
s.add(m, name='asdf')
s.add(m, parent_name='asdf')
with pytest.raises(ValueError):
s.add(m, parent_name='asfd')
with pytest.raises(TypeError):
s.add(None)
s.clear()
# Test bounds
m1 = Mesh.from_trimesh(trimesh.creation.box())
m2 = Mesh.from_trimesh(trimesh.creation.box())
m3 = Mesh.from_trimesh(trimesh.creation.box())
n1 = Node(mesh=m1)
n2 = Node(mesh=m2, translation=[1.0, 0.0, 0.0])
n3 = Node(mesh=m3, translation=[0.5, 0.0, 1.0])
s.add_node(n1)
s.add_node(n2)
s.add_node(n3)
assert np.allclose(s.bounds, [[-0.5, -0.5, -0.5], [1.5, 0.5, 1.5]])
s.clear()
s.add_node(n1)
s.add_node(n2, parent_node=n1)
s.add_node(n3, parent_node=n2)
assert np.allclose(s.bounds, [[-0.5, -0.5, -0.5], [2.0, 0.5, 1.5]])
tf = np.eye(4)
tf[:3, 3] = np.ones(3)
s.set_pose(n3, tf)
assert np.allclose(s.bounds, [[-0.5, -0.5, -0.5], [2.5, 1.5, 1.5]])
s.remove_node(n2)
assert np.allclose(s.bounds, [[-0.5, -0.5, -0.5], [0.5, 0.5, 0.5]])
s.clear()
assert np.allclose(s.bounds, 0.0)
class CanonicalScene():
def __init__(self, path_OBJ, camera_pose, min_corner, max_corner,
symmetric_transforms, viewport_size=(128, 128),
y_fov=3.14159 / 4.0, light_intensity=[0.5, 30]):
self.light_intensity = light_intensity
self.symmetric_transforms = symmetric_transforms
self.min_corner, self.max_corner = min_corner, max_corner
self.scene = Scene(bg_color=[0, 0, 0, 0])
self.light = self._build_light(light_intensity, camera_pose)
self.camera = self._build_camera(y_fov, viewport_size, camera_pose)
self.pixel_mesh = self.scene.add(color_object(path_OBJ))
self.mesh = self.scene.add(
Mesh.from_trimesh(trimesh.load(path_OBJ), smooth=True))
self.renderer = OffscreenRenderer(viewport_size[0], viewport_size[1])
self.flags_RGBA = RenderFlags.RGBA
self.flags_FLAT = RenderFlags.RGBA | RenderFlags.FLAT
def _build_light(self, light, pose):
directional_light = DirectionalLight([1.0, 1.0, 1.0], np.mean(light))
directional_light = self.scene.add(directional_light, pose=pose)
return directional_light
def _build_camera(self, y_fov, viewport_size, pose):
aspect_ratio = np.divide(*viewport_size)
camera = PerspectiveCamera(y_fov, aspectRatio=aspect_ratio)
camera = self.scene.add(camera, pose=pose)
return camera
def _sample_parameters(self, min_corner, max_corner):
mesh_transform = sample_affine_transform(min_corner, max_corner)
light_intensity = sample_uniformly(self.light_intensity)
return mesh_transform, light_intensity
def render(self):
mesh_transform, light_intensity = self._sample_parameters(
self.min_corner, self.max_corner)
mesh_rotation = mesh_transform[0:3, 0:3]
canonical_rotation = calculate_canonical_rotation(
mesh_rotation, self.symmetric_transforms)
# mesh_rotation[0:3, 0:3] = canonical_rotation
canonical_rotation = np.dot(mesh_rotation, canonical_rotation)
mesh_rotation[0:3, 0:3] = canonical_rotation
self.scene.set_pose(self.mesh, mesh_transform)
self.scene.set_pose(self.pixel_mesh, mesh_transform)
self.light.light.intensity = light_intensity
self.pixel_mesh.mesh.is_visible = False
image, depth = self.renderer.render(self.scene, self.flags_RGBA)
self.pixel_mesh.mesh.is_visible = True
image, alpha = split_alpha_channel(image)
self.mesh.mesh.is_visible = False
RGB_mask, _ = self.renderer.render(self.scene, self.flags_FLAT)
self.mesh.mesh.is_visible = True
return image, alpha, RGB_mask
def render_symmetries(self):
images, alphas, RGB_masks = [], [], []
for rotation in self.symmetric_transforms:
symmetric_transform = to_affine_matrix(rotation, np.zeros(3))
self.scene.set_pose(self.mesh, symmetric_transform)
self.scene.set_pose(self.pixel_mesh, symmetric_transform)
self.pixel_mesh.mesh.is_visible = False
image, depth = self.renderer.render(self.scene, self.flags_RGBA)
self.pixel_mesh.mesh.is_visible = True
image, alpha = split_alpha_channel(image)
self.mesh.mesh.is_visible = False
RGB_mask, _ = self.renderer.render(self.scene, self.flags_FLAT)
self.mesh.mesh.is_visible = True
images.append(image)
alphas.append(alpha)
RGB_masks.append(RGB_mask[..., 0:3])
images = np.concatenate(images, axis=1)
RGB_masks = np.concatenate(RGB_masks, axis=1)
images = np.concatenate([images, RGB_masks], axis=0)
return images
class DualView():
"""Scene that renders a single object from two different locations.
# Arguments
OBJ_filepath: String containing the path to an OBJ file.
viewport_size: List, specifying [H, W] of rendered image.
y_fov: Float indicating the vertical field of view in radians.
distance: Float. Max distance from the camera to the origin.
light: Integer representing the light intensity.
top_only: Boolean. If True images are only take from the top.
scale: Float, factor to apply to translation vector.
roll: Float, to sample [-roll, roll] rolls of the Z OpenGL camera axis.
shift: Float, to sample [-shift, shift] to move in X, Y OpenGL axes.
"""
def __init__(self,
filepath,
viewport_size=(128, 128),
y_fov=3.14159 / 4.0,
distance=0.3,
light=5.0,
top_only=True,
scale=10.0,
roll=None,
shift=None):
self._build_scene(filepath, viewport_size, light, y_fov)
self.distance, self.roll = distance, roll
self.top_only, self.shift, self.scale = top_only, shift, scale
self.renderer = OffscreenRenderer(*viewport_size)
self.RGBA = RenderFlags.RGBA
self.epsilon = 0.01
def _build_scene(self, path, size, light, y_fov):
self.scene = Scene(bg_color=[0, 0, 0, 0])
self.light = self.scene.add(DirectionalLight([1.0, 1.0, 1.0], light))
self.camera = self.scene.add(
PerspectiveCamera(y_fov, aspectRatio=np.divide(*size)))
self.mesh = self.scene.add(
Mesh.from_trimesh(trimesh.load(path), smooth=True))
self.world_origin = self.mesh.mesh.centroid
def _sample_camera_origin(self):
distance = sample_uniformly(self.distance)
camera_origin = sample_point_in_sphere(distance, self.top_only)
camera_origin = random_perturbation(camera_origin, self.epsilon)
return camera_origin
def _change_scene(self):
camera_origin = self._sample_camera_origin()
camera_to_world, world_to_camera = compute_modelview_matrices(
camera_origin, self.world_origin, self.roll, self.shift)
self.scene.set_pose(self.camera, camera_to_world)
self.scene.set_pose(self.light, camera_to_world)
return camera_to_world, world_to_camera
def render(self):
A_to_world, world_to_A = self._change_scene()
image_A, depth_A = self.renderer.render(self.scene, flags=self.RGBA)
B_to_world, world_to_B = self._change_scene()
image_B, depth_B = self.renderer.render(self.scene, flags=self.RGBA)
image_A, alpha_A = split_alpha_channel(image_A)
image_B, alpha_B = split_alpha_channel(image_B)
world_to_A = scale_translation(world_to_A, self.scale)
world_to_B = scale_translation(world_to_B, self.scale)
A_to_world = scale_translation(A_to_world, self.scale)
B_to_world = scale_translation(B_to_world, self.scale)
matrices = np.vstack([
world_to_A.flatten(),
world_to_B.flatten(),
A_to_world.flatten(),
B_to_world.flatten()
])
return {
'image_A': image_A,
'alpha_A': alpha_A,
'depth_A': depth_A,
'image_B': image_B,
'alpha_B': alpha_B,
'depth_B': depth_B,
'matrices': matrices
}