def generate_normals(self, z_batch, cam_pos, camera):
"""Generate normals from depth."""
W, H = camera['viewport'][2:]
normals = []
for z, eye in zip(z_batch, cam_pos):
camera['eye'] = eye
pcl = z_to_pcl_CC(z.squeeze(), camera)
n = self.netG2(pcl.view(H, W, 3).permute(2, 0, 1)[np.newaxis, ...])
n = n.squeeze().permute(1, 2, 0).view(-1, 3).contiguous()
normals.append(n)
return torch.stack(normals)
def test_depth_to_world_consistency(scene, batch_size):
res = render_scene(scene)
scene = make_torch_var(load_scene(scene))
pos_wc1 = res['pos'].reshape(-1, res['pos'].shape[-1])
pos_cc1 = world_to_cam(pos_wc1, None, scene['camera'])['pos']
# First test the non-batched z_to_pcl_CC method:
# NOTE: z_to_pcl_CC takes as input the Z dimension in the camera coordinate
# and gets the full (X, Y, Z) in the camera coordinate.
pos_cc2 = z_to_pcl_CC(pos_cc1[:, 2], scene['camera'])
pos_wc2 = cam_to_world(pos_cc2, None, scene['camera'])['pos']
# Test Z -> (X, Y, Z)
np.testing.assert_array_almost_equal(get_data(pos_cc1[..., :3]),
get_data(pos_cc2[..., :3]))
# Test world -> camera -> Z -> (X, Y, Z) in camera -> world
np.testing.assert_array_almost_equal(get_data(pos_wc1[..., :3]),
get_data(pos_wc2[..., :3]))
# Then test the batched version:
camera = scene['camera']
camera['eye'] = camera['eye'].repeat(batch_size, 1)
camera['at'] = camera['at'].repeat(batch_size, 1)
camera['up'] = camera['up'].repeat(batch_size, 1)
pos_wc1 = pos_wc1.repeat(batch_size, 1, 1)
pos_cc1 = world_to_cam_batched(pos_wc1, None, scene['camera'])['pos']
pos_cc2 = z_to_pcl_CC_batched(pos_cc1[..., 2], camera) # NOTE: z = -depth
pos_wc2 = cam_to_world_batched(pos_cc2, None, camera)['pos']
# Test Z -> (X, Y, Z)
np.testing.assert_array_almost_equal(get_data(pos_cc1[..., :3]),
get_data(pos_cc2[..., :3]))
# Test world -> camera -> Z -> (X, Y, Z) in camera -> world
np.testing.assert_array_almost_equal(get_data(pos_wc1[..., :3]),
get_data(pos_wc2[..., :3]))
def generate_normals(z, camera, norm_est_net_fn):
W, H = camera['viewport'][2:]
pcl = z_to_pcl_CC(z.squeeze(), camera)
n = norm_est_net_fn(pcl.view(H, W, 3).permute(2, 0, 1)[np.newaxis, ...])
return n.squeeze().permute(1, 2, 0).contiguous().view(-1, 3)