def forward_with_trimesh(self, batch):
self.load_batch(batch, include_gt=False)
with torch.no_grad():
pred_1 = self.net.depth_and_inpaint.net1.forward(self._input)
pred_abs_depth = self.net.depth_and_inpaint.get_abs_depth(
pred_1, self._input)
proj = self.net.depth_and_inpaint.proj_depth(pred_abs_depth)
pred_depth = self.net.depth_and_inpaint.base_class.postprocess(
pred_1['depth'].detach())
silhou = self.net.base_class.postprocess(self._input.silhou).detach()
pred_depth = pred_depth.cpu().numpy()
pred_depth_minmax = pred_1['depth_minmax'].detach().cpu().numpy()[0, :]
silhou = silhou.cpu().numpy()[0, 0, :, :]
pack = {'depth': pred_depth, 'depth_minmax': pred_depth_minmax}
rendered_sph = util_sph.render_spherical(pack, silhou)[None, None, ...]
rendered_sph = torch.from_numpy(rendered_sph).float().cuda()
rendered_sph = sph_pad(rendered_sph)
with torch.no_grad():
out2 = self.net.depth_and_inpaint.net2(rendered_sph)
pred_proj_sph = self.net.backproject_spherical(out2['spherical'])
pred_proj_sph = torch.transpose(pred_proj_sph, 3, 4)
pred_proj_sph = torch.flip(pred_proj_sph, [3])
proj = torch.transpose(proj, 3, 4)
proj = torch.flip(proj, [3])
refine_input = torch.cat((pred_proj_sph, proj), dim=1)
with torch.no_grad():
pred_voxel = self.net.refine_net(refine_input)
pred_1['pred_sph_full'] = out2['spherical']
pred_1['pred_sph_partial'] = rendered_sph
pred_1['pred_proj_depth'] = proj
pred_1['pred_voxel'] = pred_voxel.flip([3]).transpose(3, 4)
pred_1['pred_proj_sph_full'] = pred_proj_sph
return pred_1
def depth2sphere(data):
'''
:param data: a dictionary of an individual sample, coming from json file
:return: sph_map, sph_centered_map
'''
depth_path = data['depth_path']
#import pdb; pdb.set_trace()
depth_img = skimage.io.imread(depth_path, as_gray=True)
data['mask_img'] = np.where(depth_img > 0.001, np.ones_like(depth_img),
np.zeros_like(depth_img))
# print("my depth minmax is: ", depth_minmax)
depth_img = skimage.util.invert(depth_img)
depth_img = depth_img[None, None, :, :]
data['depth_img'] = depth_img
sph_map, sph_centered_map = render_spherical(data, res=129)
#sphere2vox(sph_centered_map)
return sph_map, sph_centered_map
def depth2sphere(data):
'''
:param data: a dictionary of an individual sample, coming from json file
:return: sph_map, sph_centered_map
'''
depth_path = data['mask'].replace('mask', 'depth').replace('.png', '1.png').replace('media', 'mnt/hamming')
mask_path = data['mask'].replace('mask', 'bin_mask').replace('media', 'mnt/hamming')
# import pdb; pdb.set_trace()
if not os.path.exists(depth_path) or not os.path.exists(mask_path): return None
# import pdb; pdb.set_trace()
depth_img = skimage.io.imread(depth_path, as_gray=True)
mask_img = skimage.io.imread(mask_path, as_gray=True)
# print("my depth minmax is: ", depth_minmax)
depth_img = skimage.util.invert(depth_img)
depth_img = depth_img[None, None, :, :]
data['depth_img'] = depth_img
data['mask_img'] = mask_img
sph_map, sph_centered_map = render_spherical(data, res=129)
return sph_map, sph_centered_map
def depth2sphere(depth_address):
mine = True
if mine:
box = True
if box:
depth = skimage.io.imread(
depth_address, as_gray=True) #[np.newaxis, np.newaxis, ...]
mask = skimage.io.imread(depth_address.replace(
"depth", "bin_mask").replace("1.png", ".png"),
as_gray=True)
with open("/mnt/Extra/dsl/GenRe-ShapeHD/gen_sphr/data/0000.json",
"r") as j_file:
data = json.load(j_file)[0]
else:
depth_address = depth_address.replace("000001-depth-001",
"000381-depth-021")
depth = skimage.io.imread(
depth_address, as_gray=True) # [np.newaxis, np.newaxis, ...]
mask = skimage.io.imread(depth_address.replace(
"depth", "bin_mask").replace("1.png", ".png"),
as_gray=True)
with open("/mnt/Extra/dsl/GenRe-ShapeHD/gen_sphr/data/0006.json",
"r") as j_file:
data = json.load(j_file)[996]
#
from scipy import ndimage as nd
# depth = nd.zoom(depth, 2, mode="nearest")
# mask = nd.zoom(mask, 2, mode="nearest")
# plt.imshow(depth)
# plt.show()
depth_minmax = data["depth_minmax"]
print("my depth minmax is: ", depth_minmax)
depth = skimage.util.invert(depth)
# depth = skimage.util.pad(depth, ([np.abs(np.subtract(*depth.shape))//2]*2, (0, 0)), 'constant', constant_values=0.)
# mask = skimage.util.pad(mask, ([np.abs(np.subtract(*mask.shape))//2]*2, (0, 0)), 'constant', constant_values=0.)
# plt.imshow(mask_img)
# plt.show()
# plt.imshow(depth)
# plt.show()
# plt.imshow(mask)
# plt.show()
# plt.imshow(depth)
# plt.show()
else:
depth_minmax = np.load(
'data/02691156_fff513f407e00e85a9ced22d91ad7027_view019.npy')
depth = skimage.io.imread('data/depth.png', as_gray=True)
mask = skimage.io.imread('data/mask.png', as_gray=True)
print("his depth minmax is: ", depth_minmax)
# plt.imshow(mask)
# plt.show()
# plt.imshow(depth)
# plt.show()
depth = depth[None, None, :, :]
# print(type(depth))
# print(depth.shape)
# There is a better way for sure. I'm not an expert in python... Save and Load was the easiest way for me...
# np.savez('/home/pvicente/software/git/GenRe-ShapeHD/tmpData/mine/tmpdata2.npz', depth=depth,
# depth_minmax=depth_minmax)
# data = np.load('/home/pvicente/software/git/GenRe-ShapeHD/tmpData/mine/tmpdata2.npz')
data = {'depth_img': depth, 'depth_minmax': depth_minmax, "mask_img": mask}
sph = render_spherical(
data, # mask,
res=129 if mine else 128
) # , I change the code in Perlis to receive the res as well, The default is 64
plt.imshow(sph)
# plt.show()
plt.savefig('depth2sphere.png')
sphere2vox(sph.astype(np.float32), res=128 if mine else 128)