def sphere2vox(sphere, res=128):
# Init
sphere = sphere.astype(np.float32)
grid = spherical_proj.gen_sph_grid(res)
proj_spherical = SphericalBackProjection().apply
# margin = 16
batch_size = 1
sph_np = sphere
sph = torch.from_numpy(sph_np.copy())
#
# print(sph.shape)
sph = sph.unsqueeze(0)
sph = sph.unsqueeze(0).cuda()
# print(sph.shape)
grid = grid.expand(1, -1, -1, -1, -1)
grid = grid[0, :, :, :, :]
grid = grid.expand(batch_size, -1, -1, -1, -1).cuda()
# crop_sph = sph[:, :, margin:h - margin, margin:w - margin]
proj_df, cnt = proj_spherical(1 - sph, grid, res)
mask = torch.clamp(cnt.detach(), 0, 1)
proj_df = (-proj_df + 1 / res) * res
proj_df = proj_df * mask
# print(proj_df.shape)
proj_df = proj_df.squeeze()
proj_df = proj_df.squeeze()
voxel = proj_df.cpu().numpy()
plot_voxel(voxel)
def __init__(self, opt, base_class):
super().__init__()
self.base_class = base_class
self.depth_and_inpaint = Depth_inpaint_net(opt, base_class)
self.refine_net = Unet_3D()
self.proj_depth = Camera_back_projection_layer()
self.joint_train = opt.joint_train
self.register_buffer('grid', gen_sph_grid())
self.grid = self.grid.expand(1, -1, -1, -1, -1)
self.proj_spherical = SphericalBackProjection().apply
self.margin = opt.padding_margin
if opt.inpaint_path is not None:
state_dicts = torch.load(opt.inpaint_path)
self.depth_and_inpaint.load_state_dict(state_dicts['nets'][0])