def test_cmp_transform_rotate_translate(self, batch_size, device, width,
height, camera_pos, object_pos,
camera_up, vertices):
mtx_rot, mtx_trans = generate_rotate_translate_matrices(
camera_pos, object_pos, camera_up)
mtx_transform = generate_transformation_matrix(camera_pos, object_pos,
camera_up)
vertices_camera = rotate_translate_points(vertices, mtx_rot, mtx_trans)
padded_vertices = torch.nn.functional.pad(vertices, (0, 1),
mode='constant',
value=1.)
vertices_camera2 = padded_vertices @ mtx_transform
assert torch.allclose(vertices_camera, vertices_camera2)
def vertices_camera(self, vertices, camera_rot, camera_trans):
return rotate_translate_points(vertices, camera_rot, camera_trans)
def test_optimize_vertex_position(self, vertices, faces, vertex_colors,
vertices_image, camera_rot, camera_trans,
camera_proj, height, width, dtype,
device):
batch_size = faces.shape[0]
# face_vertex_colors
camera_rot = camera_rot.to(device, dtype)
camera_rot.requires_grad = False
camera_trans = camera_trans.to(device, dtype)
camera_trans.requires_grad = False
camera_proj = camera_proj.to(device, dtype)
camera_proj.requires_grad = False
face_attributes = index_vertices_by_faces(
vertex_colors.to(device, dtype), faces)
vertices = vertices.to(device, dtype).clone().detach()
vertices.requires_grad = False
moved_vertices = vertices.to(device, dtype).clone()
moved_vertices[:, 0, :2] += 0.4
moved_vertices = moved_vertices.detach()
moved_vertices.requires_grad = True
images_gt = [
torch.from_numpy(
np.array(
Image.open(
os.path.join(SAMPLE_DIR, f'vertex_color_{bs}.png'))))
for bs in range(batch_size)
]
images_gt = torch.stack(images_gt, dim=0).to(device, dtype) / 255.
with torch.no_grad():
moved_vertices_camera = rotate_translate_points(
moved_vertices, camera_rot, camera_trans)
moved_vertices_image = perspective_camera(moved_vertices_camera,
camera_proj)
# test that the vertex are far enough to fail the test.
assert not torch.allclose(
moved_vertices_image, vertices_image, atol=1e-2, rtol=1e-2)
with torch.no_grad():
face_moved_vertices_camera, face_moved_vertices_image, face_moved_normals = \
prepare_vertices(moved_vertices, faces, camera_proj, camera_rot, camera_trans)
face_moved_normals_z = face_moved_normals[:, :, 2]
imfeat, _, _ = dibr_rasterization(
height, width, face_moved_vertices_camera[:, :, :, 2],
face_moved_vertices_image, face_attributes,
face_moved_normals_z)
original_loss = torch.mean(torch.abs(imfeat - images_gt))
# test that the loss is high enough
assert original_loss > 0.01
optimizer = torch.optim.Adam([moved_vertices], lr=5e-3)
for i in range(100):
optimizer.zero_grad()
face_moved_vertices_camera, face_moved_vertices_image, face_moved_normals = \
prepare_vertices(moved_vertices, faces, camera_proj, camera_rot, camera_trans)
face_moved_normals_z = face_moved_normals[:, :, 2]
imfeat, _, _ = dibr_rasterization(
height, width, face_moved_vertices_camera[:, :, :, 2],
face_moved_vertices_image, face_attributes,
face_moved_normals_z)
loss = torch.mean(torch.abs(imfeat - images_gt))
loss.backward()
optimizer.step()
moved_vertices_camera = rotate_translate_points(
moved_vertices, camera_rot, camera_trans)
moved_vertices_image = perspective_camera(moved_vertices_camera,
camera_proj)
# test that the loss went down
assert loss < 0.001
# We only test on image plan since we don't change camera angle during training we don't expect depth to be correct.
# We could probably fine-tune the test to have a lower tolerance (TODO: cfujitsang)
assert torch.allclose(moved_vertices_image,
vertices_image,
atol=1e-2,
rtol=1e-2)
