def __getitem__(self, index):
img_id = self.id_list[index]
# get sil
all_sil = np.array(
PIL.Image.open(self.dataset_dir + "sil/%08d.png" % img_id))
all_sil[all_sil < 128] = 0
all_sil[all_sil >= 128] = 1
# get parameters
with open(self.dataset_dir + "/para/%08d.json" % img_id, 'rb') as fp:
para_dic = json.load(fp)
# get src_img and pre-processing parameters
img_file = para_dic["img_file"]
src_img = np.array(PIL.Image.open(self.dataset_dir + img_file))
proc_para = para_dic["proc_para"]
# get verts and vert_norms
verts = np.array(para_dic["verts"])
vert_norms = np.array(para_dic["vert_norms"])
# get joint move and position
joint_move = np.array(para_dic["joint_move"])
joint_posi = para_dic["joint_posi"]
# get anchor move and position
achr_move = np.array(para_dic["achr_move"])
achr_posi = para_dic["achr_posi"]
# make source for joint net
sil_j = np.expand_dims(all_sil[:, :, 1], 2)
src_j = np.zeros((10, 4, 64, 64))
for i in range(len(joint_posi)):
crop_sil = center_crop(sil_j, joint_posi[i], 64)
crop_img = center_crop(src_img, joint_posi[i], 64)
crop_img = crop_img.astype(np.float)
crop_img = crop_img - crop_img[31, 31, :]
crop_img = np.absolute(crop_img)
crop_img = crop_img / 255.0
src_j[i, 0, :, :] = np.rollaxis(crop_sil, 2, 0)
src_j[i, 1:4, :, :] = np.rollaxis(crop_img, 2, 0)
# make source for anchor net
src_a = None
# commentted because prediction didn't require this
'''
sil_a = np.stack((all_sil[:,:,0], all_sil[:,:,2]), axis = -1)
src_a = np.zeros((200, 2, 32, 32))
for i in range(len(achr_posi)):
crop_sil = center_crop(sil_a, achr_posi[i], 32)
src_a[i,:,:,:] = np.rollaxis(crop_sil, 2, 0)
'''
return (src_j, src_a, src_img, joint_move, achr_move, verts,
vert_norms, proc_para, all_sil, joint_posi, achr_posi)
def __getitem__(self, index):
crt_id = self.id_list[index]
img_id = int(np.floor(crt_id / 200))
achr_id = crt_id % 200
# get sil
all_sil = np.array(PIL.Image.open(self.dataset_dir +
"sil/%08d.png" % img_id))
all_sil[all_sil<128] = 0
all_sil[all_sil>=128] = 255
# get parameters
with open (self.dataset_dir + "/para/%08d.json" % img_id, 'rb') as fp:
para_dic = json.load(fp)
achr_move = np.array(para_dic["achr_move"])
achr_posi = para_dic["achr_posi"]
# get source image
img_file = para_dic["img_file"]
src_img = np.array(PIL.Image.open(self.dataset_dir + img_file))
tgt_para = achr_move[achr_id]
tgt_para = np.expand_dims(tgt_para, 0)
if self.sil_ver is False:
src_sil = np.expand_dims(all_sil[:,:,2], 2)
crop_sil = center_crop(src_sil, achr_posi[achr_id], 32)
crop_img = center_crop(src_img, achr_posi[achr_id], 32)
crop_img = crop_img.astype(np.int)
crop_img = crop_img - crop_img[15, 15, :]
crop_img = np.absolute(crop_img)
src_in = np.concatenate((crop_sil, crop_img), axis = 2)
else:
# make input array for silhouette version
src_sil = np.stack((all_sil[:,:,0], all_sil[:,:,2]), axis = -1)
src_in = center_crop(src_sil, achr_posi[achr_id], 32)
# transform as torch tensor
src_in = PIL.Image.fromarray(src_in.astype(np.uint8))
if self.transform != None:
src_in = self.transform(src_in)
if self.get_all is True and self.sil_ver is False:
# get verts and vert_norms
verts = np.array(para_dic["verts"])
vert_norms = np.array(para_dic["vert_norms"])
proc_para = para_dic["proc_para"]
return (src_in, tgt_para, src_img, verts,
vert_norms, proc_para, all_sil)
else:
return (src_in, tgt_para)
def __getitem__(self, index):
crt_id = self.id_list[index]
img_id = int(np.floor(crt_id / 10))
joint_id = crt_id % 10
# get sil
all_sil = np.array(
PIL.Image.open(self.dataset_dir + "sil/%08d.png" % img_id))
all_sil[all_sil < 128] = 0
all_sil[all_sil >= 128] = 255
# get parameters
with open(self.dataset_dir + "/para/%08d.json" % img_id, 'rb') as fp:
para_dic = json.load(fp)
joint_move = para_dic["joint_move"]
joint_posi = para_dic["joint_posi"]
# make target para
tgt_para = np.array(joint_move[(joint_id * 2):(joint_id * 2 + 2)])
if self.sil_ver is False:
# make input array for image version
img_file = para_dic["img_file"]
src_img = np.array(PIL.Image.open(self.dataset_dir + img_file))
src_sil = np.expand_dims(all_sil[:, :, 1], 2)
crop_sil = center_crop(src_sil, joint_posi[joint_id], 64)
crop_img = center_crop(src_img, joint_posi[joint_id], 64)
crop_img = crop_img.astype(np.int)
crop_img = crop_img - crop_img[31, 31, :]
crop_img = np.absolute(crop_img)
src_in = np.concatenate((crop_sil, crop_img), axis=2)
else:
# make input array for silhouette version
src_sil = all_sil[:, :, :2]
src_in = center_crop(src_sil, joint_posi[joint_id], 64)
# transform as torch tensor
src_in = PIL.Image.fromarray(src_in.astype(np.uint8))
if self.transform != None:
src_in = self.transform(src_in)
if self.get_all is True and self.sil_ver is False:
# get verts and vert_norms
verts = np.array(para_dic["verts"])
vert_norms = np.array(para_dic["vert_norms"])
proc_para = para_dic["proc_para"]
return (src_in, tgt_para, src_img, verts, vert_norms, proc_para,
all_sil)
else:
return (src_in, tgt_para)
test_num = 0
src_img = np.array(PIL.Image.open(opt.outf + "std_img.jpg"))
hmr_mesh = om.read_trimesh(opt.outf + "hmr_mesh.obj")
hmr_mesh.request_vertex_normals()
hmr_mesh.update_normals()
verts = hmr_mesh.points()
vert_norms = hmr_mesh.vertex_normals()
# make input tensor for joint net
joint_posi = get_joint_posi(verts)
proj_sil = my_renderer.silhouette(verts=verts)
src_sil = np.expand_dims(proj_sil, 2)
src_j = np.zeros((10, 4, 64, 64))
for i in range(len(joint_posi)):
crop_sil = center_crop(src_sil, joint_posi[i], 64)
crop_img = center_crop(src_img, joint_posi[i], 64)
crop_img = crop_img.astype(np.float)
crop_img = crop_img - crop_img[31, 31, :]
crop_img = np.absolute(crop_img)
crop_img = crop_img / 255.0
src_j[i, 0, :, :] = np.rollaxis(crop_sil, 2, 0)
src_j[i, 1:4, :, :] = np.rollaxis(crop_img, 2, 0)
print("done")
# ==============================predict joint==============================
print("joint deform......", end='')
joint_tsr = pdt_j.predict_batch(src_j)
joint_para = np.array(joint_tsr.data.cpu())
joint_para = np.concatenate((joint_para, np.zeros((10, 1))), axis=1)
