def render_images(self, value):
print("Value: ", value)
if len(self.pc_codes) == 0:
return
delta = (self.pc_codes[1] - self.pc_codes[0]) / (101)
delta *= (int(value) + 1)
pc_code = self.pc_codes[0] + delta
pc = ptcloud_ae.decoder.predict(pc_code)
pc *= 0.5
target_path = os.path.join(PLOTS_PATH, value + ".png")
fig = plot_3d_point_cloud(pc[0][:, 0],
pc[0][:, 1],
pc[0][:, 2],
show=False,
azim=320,
colorize='rainbow',
filename=target_path)
if len(self.images) == 0:
image = self.show_image(target_path, x=256, y=0)
self.images.append(image)
image = render_by_pc2pix(pc_code, self.pc2pix, azim=-40)
image = self.display_image(image, x=256, y=256)
self.images.append(image)
else:
img = Image.open(target_path)
render = ImageTk.PhotoImage(img)
self.images[0].configure(image=render)
self.images[0].image = render
img = render_by_pc2pix(pc_code, self.pc2pix, azim=-40)
img = Image.fromarray(np.uint8(img * 255))
img = img.resize((256, 256), Image.ANTIALIAS)
render = ImageTk.PhotoImage(img)
self.images[1].configure(image=render)
self.images[1].image = render
def init_pc2pix(self):
# datasets = ('test')
os.makedirs(PLOTS_PATH, exist_ok=True)
# sofa2car - modify these 2 keys manually
#keys = [ "04256520", "02958343"]
# sofa2chair
keys = ["04256520", "03001627"]
# chair2chair
# keys = [ "03001627", "03001627"]
tags = []
data = js[keys[0]]
tag = data['test']
tags.append(tag)
data = js[keys[1]]
tag = data['test']
tags.append(tag)
plys = []
ply_path = os.path.join(self.ply, keys[0])
plys.append(ply_path)
ply_path = os.path.join(self.ply, keys[1])
plys.append(ply_path)
tagslen = min(len(tags[0]), len(tags[1]))
self.tags = []
self.pc_codes = []
np.random.seed(int(time.time()))
for i in range(2):
j = np.random.randint(0, tagslen, 1)[0]
tag = tags[i][j]
self.tags.append(tag)
# images = []
# pc_codes = []
ply_file = os.path.join(plys[i], tag + ".ply")
pc = load_ply(ply_file)
target_path = os.path.join(PLOTS_PATH, tag + ".png")
fig = plot_3d_point_cloud(pc[:, 0],
pc[:, 1],
pc[:, 2],
show=False,
azim=320,
colorize='rainbow',
filename=target_path)
pc = norm_pc(pc)
shape = pc.shape
pc = np.reshape(pc, [-1, shape[0], shape[1]])
pc_code = ptcloud_ae.encoder.predict(pc)
self.pc_codes.append(pc_code)
self.show_image(target_path, x=(i * 2) * 256, y=0)
image = render_by_pc2pix(pc_code, self.pc2pix, azim=-40)
print(image.shape)
self.display_image(image, x=(i * 2) * 256, y=256)
def plot_image2ptcloud_results(image2ptcloud):
im, pc, vol = image2ptcloud.test_source.next_batch()
print("pc: ", pc.shape)
save_images(im, posterior_net.test_source)
pre = image2ptcloud.predict(im)
pre *= 0.5
iou = 0.0
for i in range(len(latent)):
grd = pc[i]
print("Ground: ", i)
plot_3d_point_cloud(grd[:, 0],
grd[:, 1],
grd[:, 2],
title='point cloud',
in_u_sphere=True)
print("Prediction: ", i)
plot_3d_point_cloud(pre[i][:, 0],
pre[i][:, 1],
pre[i][:, 2],
title='point cloud',
in_u_sphere=True)
def plot_decoded_ptcloud(ae, test_source):
im, pc = test_source.next_batch()
data_dir = 'ptcloud_out'
save_dir = os.path.join(os.getcwd(), data_dir)
if not os.path.isdir(save_dir):
os.makedirs(save_dir)
for i in range(len(pc)):
test = pc[i]
plot_3d_point_cloud(test[:, 0],
test[:, 1],
test[:, 2],
in_u_sphere=True)
test = pc[i] / 0.5
shape = (1, ) + test.shape
test = np.reshape(test, shape)
reconstruction = 0.5 * ae.predict(test)
plot_3d_point_cloud(reconstruction[0][:, 0],
reconstruction[0][:, 1],
reconstruction[0][:, 2],
in_u_sphere=True)
def plot_posterior_net_results(posterior_net,
ptcloud_ae=None,
image_ae=None,
im2pc=True):
im, pc, view = posterior_net.test_source.next_batch()
print("pc: ", pc.shape)
# save_images(im, posterior_net.test_source)
if im2pc:
latent = posterior_net.posterior_net.predict(im)
pre = ptcloud_ae.decoder.predict(latent)
pre *= 0.5
cost = get_chamfer(pc, pre)
cost = np.sqrt(cost)
print("Batch chamfer: ", cost.mean())
for i in range(len(latent)):
grd = pc[i]
print("Ground: ", i)
plot_3d_point_cloud(grd[:, 0],
grd[:, 1],
grd[:, 2],
title='point cloud',
in_u_sphere=True)
print("Prediction: ", i)
print("Chamfer: ", cost[i])
plot_3d_point_cloud(pre[i][:, 0],
pre[i][:, 1],
pre[i][:, 2],
title='point cloud',
in_u_sphere=True)
else:
latent, v = posterior_net.posterior_net.predict([pc, view])
pre = image_ae.decoder.predict(latent)
for i in range(len(latent)):
grd = pc[i]
print("Prediction: ", i)
print("View ground: ", view[i], " View pred: ", v[i])
img = pre[i]
s = img.shape
img = np.reshape(img, (s[0], s[0]))
save_image(img, i, folder="predictions")
plot_3d_point_cloud(grd[:, 0],
grd[:, 1],
grd[:, 2],
title='point cloud',
in_u_sphere=True)
n_interpolate = 10
if not interpolate:
n_interpolate = 2
for i in range(tagslen - 1):
n = 0
tag = tags[i]
images = []
pc_codes = []
ply_file = os.path.join(ply_path_main, tag + ".ply")
pc = load_ply(ply_file)
target_path = os.path.join(PLOTS_PATH, tag + "_" + str(n) + ".png")
n += 1
fig = plot_3d_point_cloud(pc[:, 0],
pc[:, 1],
pc[:, 2],
show=False,
azim=320,
colorize='rainbow',
filename=target_path)
image = np.array(Image.open(target_path)) / 255.0
images.append(image)
pc = norm_pc(pc)
shape = pc.shape
pc = np.reshape(pc, [-1, shape[0], shape[1]])
pc_code1 = ptcloud_ae.encoder.predict(pc)
pc_codes.append(pc_code1)
tag = tags[i+1]
ply_file = os.path.join(ply_path_main, tag + ".ply")
pc = load_ply(ply_file)
target_path = os.path.join(PLOTS_PATH, tag + "_" + str(n_interpolate + 1) + ".png")
# elev += 360.
azim = azim[0][0]
#if azim < 0:
# azim += 360.
i = 0
t += 1
for path in paths:
target_path = os.path.join(PLOTS_PATH, tag + "-" + str(i) + ".png")
ply_file = os.path.join(path, tag + ".ply")
i += 1
ply_data = PlyData.read(ply_file)
points = ply_data['vertex']
pc = np.vstack([points['x'], points['y'], points['z']]).T
fig = plot_3d_point_cloud(pc[:, 0],
pc[:, 1],
pc[:, 2],
show=False,
elev=elev,
azim=azim,
colorize='rainbow',
filename=target_path)
#fig.close('all')
image = np.array(Image.open(target_path)) / 255.0
images.append(image)
print(str(t), view_file, ply_file, "-->", target_path, elev, azim)
plot_images(1, 5, images, tag + ".png", dir_name="point_clouds")
print(val.shape)
test = np.array(x['test'])
print(test.shape)
print(val)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
args = parser.parse_args()
load_samples()
exit(0)
dsource = DataSource()
ims, pcs = dsource.next_batch()
for i in range(len(ims)):
im = ims[i]
path = os.path.dirname(os.path.realpath(__file__))
path = os.path.join(path, "tmp")
f = "im-%d.png" % i
path = os.path.join(path, f)
dsource.save_img(im, path)
reconstruction = pcs[i]
plot_3d_point_cloud(reconstruction[:, 0],
reconstruction[:, 1],
reconstruction[:, 2],
in_u_sphere=True)
print(ims.shape)
print(pcs.shape)
dsource.close()