def get_transformed_geometry(translation, rotation, scale):
# Build bent square in object space, on z = 0 plane
vertices_object = tf.constant(
[[-1, -1, 0.], [-1, 1, 0], [1, 1, 0], [1, -1, -1.3]],
dtype=tf.float32) * square_size / 2
faces = [[0, 1, 2], [0, 2, 3]]
# ** we should add an occluding triangle!
# ** also a non-planar meeting-of-faces
vertices_object, faces = lighting.split_vertices_by_face(
vertices_object, faces)
# Convert vertices to homogeneous coordinates
vertices_object = tf.concat(
[vertices_object,
tf.ones_like(vertices_object[:, -1:])], axis=1)
# Transform vertices from object to world space, by rotating around the z-axis
vertices_world = tf.matmul(
vertices_object, matrices.rodrigues(
[0., 0., rotation])) * scale + tf.concat([translation, [0.]],
axis=0)
# Calculate face normals
normals_world = lighting.vertex_normals(vertices_world, faces)
# Transform vertices from world to camera space; note that the camera points along the negative-z axis in camera space
view_matrix = matrices.translation([-0.5, 0., -3.5
]) # translate it away from the camera
vertices_camera = tf.matmul(vertices_world, view_matrix)
# Transform vertices from camera to clip space
projection_matrix = matrices.perspective_projection(
near=0.1,
far=20.,
right=0.1,
aspect=float(canvas_height) / canvas_width)
vertices_clip = tf.matmul(vertices_camera, projection_matrix)
vertex_colours = tf.concat(
[tf.ones([3, 3]) * [0.8, 0.5, 0.],
tf.ones([3, 3]) * [0.5, 0.8, 0.]],
axis=0)
return vertices_clip, faces, normals_world, vertex_colours
def main():
# Build the scene geometry, which is just an axis-aligned cube centred at the origin in world space
cube_vertices_object = []
cube_uvs = []
cube_faces = []
def add_quad(vertices, uvs):
index = len(cube_vertices_object)
cube_faces.extend([[index + 2, index + 1, index],
[index, index + 3, index + 2]])
cube_vertices_object.extend(vertices)
cube_uvs.extend(uvs)
add_quad(vertices=[[-1, -1, 1], [1, -1, 1], [1, 1, 1], [-1, 1, 1]],
uvs=[[0.1, 0.9], [0.9, 0.9], [0.9, 0.1], [0.1, 0.1]]) # front
add_quad(vertices=[[-1, -1, -1], [1, -1, -1], [1, 1, -1], [-1, 1, -1]],
uvs=[[1, 1], [0, 1], [0, 0], [1, 0]]) # back
add_quad(vertices=[[1, 1, 1], [1, 1, -1], [1, -1, -1], [1, -1, 1]],
uvs=[[0.3, 0.25], [0.6, 0.25], [0.6, 0.55], [0.3, 0.55]]) # right
add_quad(vertices=[[-1, 1, 1], [-1, 1, -1], [-1, -1, -1], [-1, -1, 1]],
uvs=[[0.4, 0.4], [0.5, 0.4], [0.5, 0.5], [0.4, 0.5]]) # left
add_quad(vertices=[[-1, 1, -1], [1, 1, -1], [1, 1, 1], [-1, 1, 1]],
uvs=[[0, 0], [2, 0], [2, 2], [0, 2]]) # top
add_quad(vertices=[[-1, -1, -1], [1, -1, -1], [1, -1, 1], [-1, -1, 1]],
uvs=[[0, 0], [2, 0], [2, 2], [0, 2]]) # bottom
cube_vertices_object = np.asarray(cube_vertices_object, np.float32)
cube_uvs = np.asarray(cube_uvs, np.float32)
# Load the texture image
texture = tf.cast(
tf.image.decode_jpeg(
tf.read_file(os.path.dirname(__file__) + '/cat.jpg')),
tf.float32) / 255.
# Convert vertices to homogeneous coordinates
cube_vertices_object = tf.concat(
[cube_vertices_object,
tf.ones_like(cube_vertices_object[:, -1:])],
axis=1)
# Transform vertices from object to world space, by rotating around the vertical axis
cube_vertices_world = tf.matmul(cube_vertices_object,
matrices.rodrigues([0., 0.6, 0.]))
# Calculate face normals
cube_normals_world = lighting.vertex_normals(cube_vertices_world,
cube_faces)
# Transform vertices from world to camera space; note that the camera points along the negative-z axis in camera space
view_matrix = matrices.compose(
matrices.translation([0., -2.,
-3.2]), # translate it away from the camera
matrices.rodrigues([-0.5, 0., 0.]) # tilt the view downwards
)
cube_vertices_camera = tf.matmul(cube_vertices_world, view_matrix)
# Transform vertices from camera to clip space
projection_matrix = matrices.perspective_projection(
near=0.1, far=20., right=0.1, aspect=float(frame_height) / frame_width)
cube_vertices_clip = tf.matmul(cube_vertices_camera, projection_matrix)
# The following function is applied to the G-buffer, which is a multi-channel image containing all the vertex attributes. It
# uses this to calculate the shading (texture and lighting) at each pixel, hence their final intensities
def shader_fn(gbuffer, texture, light_direction):
# Unpack the different attributes from the G-buffer
mask = gbuffer[:, :, :1]
uvs = gbuffer[:, :, 1:3]
normals = gbuffer[:, :, 3:]
# Sample the texture at locations corresponding to each pixel; this defines the unlit material color at each point
unlit_colors = sample_texture(
texture, uvs_to_pixel_indices(uvs,
tf.shape(texture)[:2]))
# Calculate a simple grey ambient lighting component
ambient_contribution = unlit_colors * [0.4, 0.4, 0.4]
# Calculate a diffuse (Lambertian) lighting component
diffuse_contribution = lighting.diffuse_directional(
tf.reshape(normals, [-1, 3]),
tf.reshape(unlit_colors, [-1, 3]),
light_direction,
light_color=[0.6, 0.6, 0.6],
double_sided=True)
diffuse_contribution = tf.reshape(diffuse_contribution,
[frame_height, frame_width, 3])
# The final pixel intensities inside the shape are given by combining the ambient and diffuse components;
# outside the shape, they are set to a uniform background color
pixels = (diffuse_contribution +
ambient_contribution) * mask + [0., 0., 0.3] * (1. - mask)
return pixels
# Render the G-buffer channels (mask, UVs, and normals at each pixel), then perform the deferred shading calculation
# In general, any tensor required by shader_fn and wrt which we need derivatives should be included in shader_additional_inputs;
# although in this example they are constant, we pass the texture and lighting direction through this route as an illustration
light_direction = tf.linalg.l2_normalize([1., -0.3, -0.5])
pixels = dirt.rasterise_deferred(
vertices=cube_vertices_clip,
vertex_attributes=tf.concat(
[
tf.ones_like(cube_vertices_object[:, :1]), # mask
cube_uvs, # texture coordinates
cube_normals_world # normals
],
axis=1),
faces=cube_faces,
background_attributes=tf.zeros([frame_height, frame_width, 6]),
shader_fn=shader_fn,
shader_additional_inputs=[texture, light_direction])
save_pixels = tf.write_file(
'textured.jpg', tf.image.encode_jpeg(tf.cast(pixels * 255, tf.uint8)))
session = tf.Session(config=tf.ConfigProto(gpu_options=tf.GPUOptions(
allow_growth=True)))
with session.as_default():
save_pixels.run()
n_tex_para = 40
mat_data = sio.loadmat(mat_filename)
shape_param = tf.constant(mat_data['Shape_Para'], dtype=tf.float32)
shape_param = tf.expand_dims(shape_param, 0)
exp_param = tf.constant(mat_data['Exp_Para'], dtype=tf.float32)
exp_param = tf.expand_dims(exp_param, 0)
tex_param = tf.constant(mat_data['Tex_Para'][:n_tex_para, :],
dtype=tf.float32)
tex_param = tf.expand_dims(tex_param, 0)
color_param = tf.constant(mat_data['Color_Para'], dtype=tf.float32)
color_param = tf.expand_dims(color_param, 0)
illum_param = tf.constant(mat_data['Illum_Para'], dtype=tf.float32)
illum_param = tf.expand_dims(illum_param, 0)
pose_param = tf.constant(mat_data['Pose_Para'], dtype=tf.float32)
pose_param = tf.expand_dims(pose_param, 0)
tf_bfm = TfMorphableModel(model_path='../../examples/Data/BFM/Out/BFM.mat',
n_tex_para=n_tex_para)
vertices = tf_bfm.get_vertices(shape_param=shape_param,
exp_param=exp_param,
batch_size=1)
from dirt import lighting
vertex_norm = lighting.vertex_normals(vertices, tf_bfm.triangles)
texture = tf_bfm.get_vertex_colors(tex_param, color_param, illum_param,
-vertex_norm, 1)
def dirt_rendering_orthographic( vertices, faces, reflectances, \
frame_width, frame_height, cx, cy, vertices_to_keep_track = None ):
if vertices.shape[-1] == 2:
vertices = tf.concat(
[vertices, tf.zeros_like(vertices[..., -1:])], axis=-1)
vertical_flip = False
if frame_height < 0:
frame_height = -frame_height
vertical_flip = True
if reflectances is None:
reflectances = tf.ones_like(vertices)
rendering_batch_size = tf.shape(vertices)[0]
ones = tf.ones([rendering_batch_size], dtype=tf.float32)
normals = lighting.vertex_normals(vertices, faces)
bz_facemodel_vertices_object, bz_facemodel_vertex_colors, bz_facemodel_vertex_normals, bz_facemodel_faces = split_vertices_by_color_normal_face(
vertices, reflectances, normals, faces)
if vertices_to_keep_track is None:
landmark_vertices = vertices
landmark_normals = normals
else:
landmark_vertices, landmark_normals = get_landmarks(
vertices, normals, vertices_to_keep_track)
translation_parameters = np.array([[0, 0, 100]], dtype=np.float32)
translation_parameters = tf.tile(
tf.constant(translation_parameters, dtype=tf.float32),
[rendering_batch_size, 1])
rotation_parameters = np.array([[1e-10, 0, 0]], dtype=np.float32)
rotation_parameters = tf.tile(
tf.constant(rotation_parameters, dtype=tf.float32),
[rendering_batch_size, 1])
cx = tf.ones([rendering_batch_size], dtype=tf.float32) * cx
cy = tf.ones([rendering_batch_size],
dtype=tf.float32) * (frame_height - cy)
# Transform vertices from world to camera space; note that the camera points along the negative-z axis in camera space
view_matrix = matrices.compose(
matrices.translation(
translation_parameters), # translate it away from the camera
matrices.rodrigues(rotation_parameters) # tilt the view downwards
)
"""
with tf.Session() as sess:
_batch = sess.run( rendering_batch_size )
_a = sess.run( matrices.translation(translation_parameters) )
_b = sess.run( matrices.rodrigues(rotation_parameters) )
_c = sess.run( view_matrix )
print("translation", np.transpose(_a[0]))
print("rotation", np.transpose(_b[0]))
print("camera", np.transpose(_c[0]))
"""
# Convert vertices to homogeneous coordinates
bz_facemodel_vertices_object = tf.concat([
bz_facemodel_vertices_object,
tf.ones_like(bz_facemodel_vertices_object[..., -1:])
],
axis=-1)
landmark_vertices = tf.concat(
[landmark_vertices,
tf.ones_like(landmark_vertices[..., -1:])],
axis=-1)
viewpoint_direction = -tf.ones_like(translation_parameters)
norms = tf.norm(viewpoint_direction, axis=-1,
keep_dims=True) # indexed by *, singleton
viewpoint_direction /= norms
viewpoint_direction = np.array([0, 0, -1], np.float32)
viewpoint_direction = np.expand_dims(viewpoint_direction, axis=0)
viewpoint_direction = tf.constant(viewpoint_direction)
viewpoint_direction = tf.tile(viewpoint_direction,
[rendering_batch_size, 1])
# Transform vertices from object to world space, by rotating around the vertical axis
bz_facemodel_vertices_world = bz_facemodel_vertices_object # tf.matmul(cube_vertices_object, [matrices.rodrigues([0.0, 0.0, 0.0])] )
landmark_vertices_world = landmark_vertices
# Calculate face normals; pre_split implies that no faces share a vertex
bz_facemodel_faces = tf.expand_dims(bz_facemodel_faces, axis=0)
bz_facemodel_faces = tf.tile(bz_facemodel_faces,
(rendering_batch_size, 1, 1))
bz_facemodel_normals_world = bz_facemodel_vertex_normals # lighting.vertex_normals_pre_split(cube_vertices_world, cube_faces)
bz_facemodel_vertices_camera = tf.matmul(bz_facemodel_vertices_world,
view_matrix)
# Transform vertices from camera to clip space
near = 10.0 * ones
far = 200.0 * ones
projection_matrix = orthographic_projection(near=near,
far=far,
w=frame_width,
ones=ones,
h=frame_height,
cx=cx,
cy=cy)
bz_facemodel_vertices_clip = tf.matmul(bz_facemodel_vertices_camera,
projection_matrix)
landmark_vertices_vertices_clip = tf.matmul(landmark_vertices_world,
projection_matrix)
"""
with tf.Session() as sess:
_v0 = sess.run( bz_facemodel_vertices_world )
_v1 = sess.run( bz_facemodel_vertices_camera )
_v2 = sess.run( bz_facemodel_vertices_clip )
print("vertices\n", (_v0[0]))
print("vertices\n", (_v1[0]))
print("vertices\n", (_v2[0]))
"""
full_model_pixels = dirt.rasterise_batch(
vertices=bz_facemodel_vertices_clip,
faces=bz_facemodel_faces,
vertex_colors=bz_facemodel_vertex_colors,
background=tf.zeros(
[rendering_batch_size, frame_height, frame_width, 3]),
width=frame_width,
height=frame_height,
channels=3)
# landmarks
denom = landmark_vertices_vertices_clip[:, :, 3]
landmark_points = landmark_vertices_vertices_clip[:, :, 0:2] / (
denom[..., tf.newaxis])
landmark_xpoint = (1.0 + landmark_points[:, :, 0:1]) * frame_width * 0.5
landmark_ypoint = (1.0 + landmark_points[:, :, 1:2]) * frame_height * 0.5
landmark_points = tf.concat([landmark_xpoint, landmark_ypoint], axis=-1)
output_dictionary = {}
output_dictionary["rendering_results"] = full_model_pixels
output_dictionary["vertices"] = landmark_points
output_dictionary["landmark_normals"] = landmark_normals
"""
with tf.Session() as sess:
_vertices = sess.run( bz_facemodel_vertices_clip )
print( _vertices.shape )
print( _vertices )
"""
if vertical_flip is True:
output_dictionary["rendering_results"] = output_dictionary[
"rendering_results"][:, ::-1, :, :]
return output_dictionary
def dirt_rendering( vertices, faces, reflectances, \
spherical_harmonics_parameters, translation_parameters, camera_rotation_parameters, base_rotation, rotation_parameters, \
frame_width, frame_height, focal_length, cx, cy, vertices_indices ):
rendering_batch_size = tf.shape(vertices)[0]
vertices = tf.matmul(vertices,
matrices.rodrigues(rotation_parameters)[..., :3, :3])
vertex_normals = normals = lighting.vertex_normals(vertices, faces)
landmark_vertices, landmark_normals = get_landmarks(
vertices, normals, vertices_indices)
bz_facemodel_vertices_object, bz_facemodel_vertex_colors, bz_facemodel_vertex_normals, bz_facemodel_faces = split_vertices_by_color_normal_face(
vertices, reflectances, normals, faces)
# Transform vertices from world to camera space; note that the camera points along the negative-z axis in camera space
view_matrix = matrices.compose(
# matrices.rodrigues(rotation_parameters),
matrices.translation(-translation_parameters
), # translate it away from the camera
matrices.rodrigues(
camera_rotation_parameters), # tilt the view downwards
matrices.rodrigues(base_rotation))
"""
view_matrix = matrices.compose(
matrices.translation(translation_parameters), # translate it away from the camera
matrices.rodrigues(camera_rotation_parameters), # tilt the view downwards
matrices.translation(-translation_parameters) # translate it away from the camera
)
"""
# Convert vertices to homogeneous coordinates
bz_facemodel_vertices_object = tf.concat([
bz_facemodel_vertices_object,
tf.ones_like(bz_facemodel_vertices_object[..., -1:])
],
axis=-1)
landmark_vertices = tf.concat(
[landmark_vertices,
tf.ones_like(landmark_vertices[..., -1:])],
axis=-1)
"""
with tf.Session() as sess:
_view_matrix = sess.run( view_matrix )
print( _view_matrix )
"""
viewpoint_direction = -tf.ones_like(translation_parameters)
norms = tf.norm(viewpoint_direction, axis=-1,
keep_dims=True) # indexed by *, singleton
viewpoint_direction /= (norms + 1e-5)
viewpoint_direction = np.array([0, 0, 1], np.float32)
viewpoint_direction = np.expand_dims(viewpoint_direction, axis=0)
viewpoint_direction = tf.constant(viewpoint_direction)
viewpoint_direction = tf.tile(viewpoint_direction,
[rendering_batch_size, 1])
# Transform vertices from object to world space, by rotating around the vertical axis
bz_facemodel_vertices_world = bz_facemodel_vertices_object # tf.matmul(cube_vertices_object, [matrices.rodrigues([0.0, 0.0, 0.0])] )
landmark_vertices_world = landmark_vertices
# Calculate face normals; pre_split implies that no faces share a vertex
bz_facemodel_faces = tf.expand_dims(bz_facemodel_faces, axis=0)
bz_facemodel_faces = tf.tile(bz_facemodel_faces,
(rendering_batch_size, 1, 1))
bz_facemodel_normals_world = bz_facemodel_vertex_normals # lighting.vertex_normals_pre_split(cube_vertices_world, cube_faces)
landmark_vertices_normals = landmark_normals
bz_facemodel_vertices_camera = tf.matmul(bz_facemodel_vertices_world,
view_matrix)
landmark_vertices_camera = tf.matmul(landmark_vertices_world, view_matrix)
# Transform vertices from camera to clip space
near = focal_length * 0.1 # 0.01 just constant
far = focal_length * 100
# right = frame_width * 0.5 * 0.01
# projection_matrix = matrices.perspective_projection(near=near, far=1000., right=right, aspect=float(frame_height) / frame_width)
projection_matrix = matrices.perspective_projection( near=near, far=far, fx = focal_length, fy=focal_length, \
w = frame_width, h = frame_height, cx = cx, cy = cy )
# projection_matrix = tf.expand_dims( projection_matrix, axis = 0)
# projection_matrix = tf.tile( projection_matrix, (rendering_batch_size,1,1))
bz_facemodel_vertices_clip = tf.matmul(bz_facemodel_vertices_camera,
projection_matrix)
landmark_vertices_vertices_clip = tf.matmul(landmark_vertices_camera,
projection_matrix)
#K = tf.constant(
# np.array( [ [focal_length,0,frame_width*0.5],[0,focal_length,frame_height*0.5],[0.0,0.0,1.0]] ), dtype = tf.float32)
"""
### if you want check projection matrix
with tf.Session() as sess:
_landmarks = sess.run( landmark_vertices )
_view_matrix = sess.run( view_matrix )
_projection_matrix = sess.run( projection_matrix )
#_K = sess.run(K )
_T = sess.run( matrices.translation(translation_parameters) )
_R = sess.run( matrices.rodrigues(rotation_parameters) )
print(np.array2string(_T, separator=', '))
print(np.array2string(_R, separator=', '))
print(np.array2string(_landmarks, separator=', '))
print(np.array2string(_view_matrix, separator=', '))
#print(np.array2string(_K, separator=', '))
print(np.array2string(_projection_matrix, separator=', '))
"""
"""
# Calculate lighting, as combination of diffuse and ambient
vertex_colors_lit = diffuse_directional_lights(
bz_facemodel_normals_world, bz_facemodel_vertex_colors,
light_direction=light_directions, viewpoint_direction=viewpoint_direction, light_color=light_colors
) * 1.0 #+ bz_facemodel_vertex_colors * 0.2
"""
# geometry
use_spherical_harmonics = False
if use_spherical_harmonics == True:
vertex_colors_lit = spherical_harmonics(
bz_facemodel_normals_world,
bz_facemodel_vertex_colors,
spherical_harmonics_parameters,
viewpoint_direction=viewpoint_direction)
geometry_visualization_spherical_harmonics_parameters = np.zeros(
[27], dtype=np.float32)
geometry_visualization_spherical_harmonics_parameters[3:3 + 9] = 1.0
geometry_visualization_spherical_harmonics_parameters = tf.constant(
geometry_visualization_spherical_harmonics_parameters,
dtype=tf.float32)
geometry_visualization_vertex_colors_lit = spherical_harmonics( bz_facemodel_normals_world, tf.ones_like(bz_facemodel_vertex_colors), \
geometry_visualization_spherical_harmonics_parameters, viewpoint_direction=viewpoint_direction )
else:
light_directions = np.array([[0, 0, 1]])
norm = np.linalg.norm(light_directions, axis=-1)
light_directions = light_directions / norm[:, np.newaxis]
light_directions = tf.constant(light_directions, dtype=tf.float32)
light_directions = tf.tile(light_directions, (rendering_batch_size, 1))
light_colors = tf.constant([1., 1., 1.], dtype=tf.float32)
light_colors = tf.expand_dims(light_colors, axis=0)
light_colors = tf.tile(light_colors, (rendering_batch_size, 1))
geometry_visualization_vertex_colors_lit = diffuse_directional_lights( bz_facemodel_normals_world, tf.ones_like(bz_facemodel_vertex_colors), \
light_direction=light_directions, viewpoint_direction=viewpoint_direction, light_color=light_colors, double_sided = False ) * 1.0 #+ bz_facemodel_vertex_colors * 0.2
vertex_colors_lit = diffuse_directional_lights( bz_facemodel_normals_world, bz_facemodel_vertex_colors, light_direction=light_directions, \
viewpoint_direction=viewpoint_direction, light_color=light_colors ) * 1.0 #+ bz_facemodel_vertex_colors * 0.2
# reflectance
reflectance_visualization_spherical_harmonics_parameters = np.zeros(
[27], dtype=np.float32)
reflectance_visualization_spherical_harmonics_parameters[0:3] = 3.0
reflectance_visualization_spherical_harmonics_parameters = tf.constant(
reflectance_visualization_spherical_harmonics_parameters,
dtype=tf.float32)
reflectance_visualization_vertex_colors_lit = spherical_harmonics( bz_facemodel_normals_world, bz_facemodel_vertex_colors, \
reflectance_visualization_spherical_harmonics_parameters, viewpoint_direction=viewpoint_direction )
# illumination
illumination_visualization_vertex_colors_lit = spherical_harmonics( bz_facemodel_normals_world, tf.ones_like(bz_facemodel_vertex_colors), \
spherical_harmonics_parameters, viewpoint_direction=viewpoint_direction )
# depth
depth_vertex_colors_lit = tf.concat([
geometry_visualization_vertex_colors_lit[:, :, 0:1],
bz_facemodel_normals_world[:, :, 2:3],
bz_facemodel_vertices_camera[:, :, 2:3]
],
axis=-1)
# landmarks
denom = landmark_vertices_vertices_clip[:, :, 3]
landmark_points = landmark_vertices_vertices_clip[:, :, 0:2] / (
denom[..., tf.newaxis])
landmark_xpoint = (1.0 + landmark_points[:, :, 0:1]) * frame_width * 0.5
landmark_ypoint = (1.0 - landmark_points[:, :, 1:2]) * frame_height * 0.5
landmark_points = tf.concat([landmark_xpoint, landmark_ypoint], axis=-1)
"""
with tf.Session() as sess:
_AAA = sess.run( landmark_points )
print(_AAA)
"""
# landmark visibility
landmark_visibility = tf.matmul(landmark_vertices_normals,
viewpoint_direction[..., tf.newaxis])
#visibility = tf.minimum( landmark_visibility+0.9999, 1.0)
#visibility = tf.cast( visibility, tf.int32 )
full_model_pixels = dirt.rasterise_batch(
vertices=bz_facemodel_vertices_clip,
faces=bz_facemodel_faces,
vertex_colors=vertex_colors_lit,
background=tf.zeros(
[rendering_batch_size, frame_height, frame_width, 3]),
width=frame_width,
height=frame_height,
channels=3)
geometry_model_pixels = dirt.rasterise_batch(
vertices=bz_facemodel_vertices_clip,
faces=bz_facemodel_faces,
vertex_colors=geometry_visualization_vertex_colors_lit,
background=tf.zeros(
[rendering_batch_size, frame_height, frame_width, 3]),
width=frame_width,
height=frame_height,
channels=3)
reflectance_model_pixels = dirt.rasterise_batch(
vertices=bz_facemodel_vertices_clip,
faces=bz_facemodel_faces,
vertex_colors=tf.ones_like(
reflectance_visualization_vertex_colors_lit),
background=tf.zeros(
[rendering_batch_size, frame_height, frame_width, 3]),
width=frame_width,
height=frame_height,
channels=3)
illumination_model_pixels = dirt.rasterise_batch(
vertices=bz_facemodel_vertices_clip,
faces=bz_facemodel_faces,
vertex_colors=illumination_visualization_vertex_colors_lit,
background=tf.zeros(
[rendering_batch_size, frame_height, frame_width, 3]),
width=frame_width,
height=frame_height,
channels=3)
depth_maps = -dirt.rasterise_batch(
vertices=bz_facemodel_vertices_clip,
faces=bz_facemodel_faces,
vertex_colors=depth_vertex_colors_lit,
background=tf.zeros(
[rendering_batch_size, frame_height, frame_width, 3]),
width=frame_width,
height=frame_height,
channels=3)
output1_dict = {}
output1_dict["full_model_pixels"] = full_model_pixels
output1_dict["vertices_clip"] = bz_facemodel_vertices_clip
output1_dict["landmark_points"] = landmark_points
output1_dict["landmark_visibility"] = landmark_visibility
#output1_dict["depth_masks"] = tf.clip_by_value( 5*( -depth_maps[:,:,:,0] ), 0, 1 ) # tf.stop_gradient( ???
#output1_dict["depth_masks"] = tf.stop_gradient( tf.clip_by_value( 100*( -depth_maps[:,:,:,0] ), 0, 1 ) )
output1_dict["depth_maps"] = depth_maps[:, :, :, 2]
output1_dict["geometry_model_pixels"] = geometry_model_pixels
output1_dict["reflectance_model_pixels"] = reflectance_model_pixels
output1_dict["illumination_model_pixels"] = illumination_model_pixels
output1_dict["surface_normals"] = (1 + depth_maps[:, :, :, 1]) / 2.0
output1_dict["vertex_normals"] = vertex_normals
return output1_dict
def main():
# Build the scene geometry, which is just an axis-aligned cube centred at the origin in world space
cube_vertices_object = []
cube_uvs = []
cube_faces = []
def add_quad(vertices, uvs):
index = len(cube_vertices_object)
cube_faces.extend([[index + 2, index + 1, index],
[index, index + 3, index + 2]])
cube_vertices_object.extend(vertices)
cube_uvs.extend(uvs)
add_quad(vertices=[[-1, -1, 1], [1, -1, 1], [1, 1, 1], [-1, 1, 1]],
uvs=[[0.1, 0.9], [0.9, 0.9], [0.9, 0.1], [0.1, 0.1]]) # front
add_quad(vertices=[[-1, -1, -1], [1, -1, -1], [1, 1, -1], [-1, 1, -1]],
uvs=[[1, 1], [0, 1], [0, 0], [1, 0]]) # back
add_quad(vertices=[[1, 1, 1], [1, 1, -1], [1, -1, -1], [1, -1, 1]],
uvs=[[0.4, 0.35], [0.5, 0.35], [0.5, 0.45], [0.4, 0.45]]) # right
add_quad(vertices=[[-1, 1, 1], [-1, 1, -1], [-1, -1, -1], [-1, -1, 1]],
uvs=[[0.4, 0.4], [0.5, 0.4], [0.5, 0.5], [0.4, 0.5]]) # left
add_quad(vertices=[[-1, 1, -1], [1, 1, -1], [1, 1, 1], [-1, 1, 1]],
uvs=[[0, 0], [2, 0], [2, 2], [0, 2]]) # top
add_quad(vertices=[[-1, -1, -1], [1, -1, -1], [1, -1, 1], [-1, -1, 1]],
uvs=[[0, 0], [2, 0], [2, 2], [0, 2]]) # bottom
cube_vertices_object = np.asarray(cube_vertices_object, np.float32)
cube_uvs = np.asarray(cube_uvs, np.float32)
# Load the texture image
texture = tf.constant(
imageio.imread(os.path.dirname(__file__) + '/cat.jpg'),
dtype=tf.float32) / 255.
# Convert vertices to homogeneous coordinates
cube_vertices_object = tf.concat(
[cube_vertices_object,
tf.ones_like(cube_vertices_object[:, -1:])],
axis=1)
# Transform vertices from object to world space, by rotating around the vertical axis
cube_vertices_world = tf.matmul(cube_vertices_object,
matrices.rodrigues([0., 0.6, 0.]))
# Calculate face normals
cube_normals_world = lighting.vertex_normals(cube_vertices_world,
cube_faces)
# Transform vertices from world to camera space; note that the camera points along the negative-z axis in camera space
view_matrix = matrices.compose(
matrices.translation([0., -2.,
-3.2]), # translate it away from the camera
matrices.rodrigues([-0.5, 0., 0.]) # tilt the view downwards
)
cube_vertices_camera = tf.matmul(cube_vertices_world, view_matrix)
# Transform vertices from camera to clip space
projection_matrix = matrices.perspective_projection(
near=0.1, far=20., right=0.1, aspect=float(frame_height) / frame_width)
cube_vertices_clip = tf.matmul(cube_vertices_camera, projection_matrix)
# Render the G-buffer channels (mask, UVs, and normals at each pixel) needed for deferred shading
gbuffer_mask = dirt.rasterise(
vertices=cube_vertices_clip,
faces=cube_faces,
vertex_colors=tf.ones_like(cube_vertices_object[:, :1]),
background=tf.zeros([frame_height, frame_width, 1]),
width=frame_width,
height=frame_height,
channels=1)[..., 0]
background_value = -1.e4
gbuffer_vertex_uvs = dirt.rasterise(
vertices=cube_vertices_clip,
faces=cube_faces,
vertex_colors=tf.concat(
[cube_uvs, tf.zeros_like(cube_uvs[:, :1])], axis=1),
background=tf.ones([frame_height, frame_width, 3]) * background_value,
width=frame_width,
height=frame_height,
channels=3)[..., :2]
gbuffer_vertex_normals_world = dirt.rasterise(
vertices=cube_vertices_clip,
faces=cube_faces,
vertex_colors=cube_normals_world,
background=tf.ones([frame_height, frame_width, 3]) * background_value,
width=frame_width,
height=frame_height,
channels=3)
# Dilate the normals and UVs to ensure correct gradients on the silhouette
gbuffer_mask = gbuffer_mask[:, :, None]
gbuffer_vertex_normals_world_dilated = tf.nn.max_pool(
gbuffer_vertex_normals_world[None, ...],
ksize=[1, 3, 3, 1],
strides=[1, 1, 1, 1],
padding='SAME')[0]
gbuffer_vertex_normals_world = gbuffer_vertex_normals_world * gbuffer_mask + gbuffer_vertex_normals_world_dilated * (
1. - gbuffer_mask)
gbuffer_vertex_uvs_dilated = tf.nn.max_pool(gbuffer_vertex_uvs[None, ...],
ksize=[1, 3, 3, 1],
strides=[1, 1, 1, 1],
padding='SAME')[0]
gbuffer_vertex_uvs = gbuffer_vertex_uvs * gbuffer_mask + gbuffer_vertex_uvs_dilated * (
1. - gbuffer_mask)
# Calculate the colour buffer, by sampling the texture according to the rasterised UVs
gbuffer_colours = gbuffer_mask * sample_texture(
texture, uvs_to_pixel_indices(gbuffer_vertex_uvs,
tf.shape(texture)[:2]))
# Calculate a simple grey ambient lighting component
ambient_contribution = gbuffer_colours * [0.4, 0.4, 0.4]
# Calculate a diffuse (Lambertian) lighting component
light_direction = unit([1., -0.3, -0.5])
diffuse_contribution = lighting.diffuse_directional(
tf.reshape(gbuffer_vertex_normals_world, [-1, 3]),
tf.reshape(gbuffer_colours, [-1, 3]),
light_direction,
light_color=[0.6, 0.6, 0.6],
double_sided=True)
diffuse_contribution = tf.reshape(diffuse_contribution,
[frame_height, frame_width, 3])
# Final pixels are given by combining the ambient and diffuse components
pixels = diffuse_contribution + ambient_contribution
session = tf.Session(config=tf.ConfigProto(gpu_options=tf.GPUOptions(
allow_growth=True)))
with session.as_default():
pixels_eval = pixels.eval()
imageio.imsave('textured.jpg', (pixels_eval * 255).astype(np.uint8))
def get_vertex_normals(vertices, faces):
""" Get vertex normals. [Vtx, Face_def] -> [Vtx_normals] """
vtx_nrm = lighting.vertex_normals(vertices=vertices, faces=faces)
return vtx_nrm