def visu_network_input_pred(self,
tag,
epoch,
data,
images,
target,
cam,
max_images=10,
store=False,
jupyter=False,
method='def'):
num = min(max_images, data.shape[0])
fig = plt.figure(figsize=(10.5, num * 3.5))
for i in range(num):
# real render input
n_render = f'batch{i}_render.png'
n_real = f'batch{i}_real.png'
real = np.transpose(
data[i, :3, :, :].cpu().numpy().astype(np.uint8), (1, 2, 0))
render = np.transpose(
data[i, 3:, :, :].cpu().numpy().astype(np.uint8), (1, 2, 0))
fig.add_subplot(num, 3, i * 3 + 1)
plt.imshow(real)
plt.tight_layout()
fig.add_subplot(num, 3, i * 3 + 2)
plt.imshow(render)
plt.tight_layout()
# prediction
masked_idx = backproject_points(target[i],
fx=cam[i, 2],
fy=cam[i, 3],
cx=cam[i, 0],
cy=cam[i, 1])
for j in range(masked_idx.shape[0]):
try:
images[i,
int(masked_idx[j, 0]),
int(masked_idx[j, 1]), 0] = 0
images[i,
int(masked_idx[j, 0]),
int(masked_idx[j, 1]), 1] = 255
images[i,
int(masked_idx[j, 0]),
int(masked_idx[j, 1]), 2] = 0
except:
pass
min1 = torch.min(masked_idx[:, 0])
max1 = torch.max(masked_idx[:, 0])
max2 = torch.max(masked_idx[:, 1])
min2 = torch.min(masked_idx[:, 1])
bb = BoundingBox(p1=torch.stack([min1, min2]),
p2=torch.stack([max1, max2]))
bb_img = bb.plot(images[i, :, :, :3].cpu().numpy().astype(
np.uint8))
fig.add_subplot(num, 3, i * 3 + 3)
plt.imshow(bb_img)
# fig.add_subplot(num, 2, i * 2 + 4)
# real = images[i, :, :, :3].cpu().numpy().astype(np.uint8)
# plt.imshow(real)
if method != 'def':
a = get_img_from_fig(fig).astype(np.uint8)
plt.close()
return a
if store:
#store_ar = (img_d* 255).round().astype(np.uint8)
plt.savefig(
f'{self.p_visu}/{str(epoch)}_{tag}_network_input_and_prediction.png',
dpi=300)
#save_image(img_d, tag=str(epoch) + tag, p_store=self.p_visu)
if jupyter:
plt.show()
if self.writer is not None:
# you can get a high-resolution image as numpy array!!
plot_img_np = get_img_from_fig(fig)
self.writer.add_image(tag,
plot_img_np,
global_step=epoch,
dataformats='HWC')
plt.close()
def plot_batch_projection(self,
tag,
epoch,
images,
target,
cam,
max_images=10,
store=False,
jupyter=False,
method='def'):
num = min(max_images, target.shape[0])
fig = plt.figure(figsize=(7, num * 3.5))
for i in range(num):
masked_idx = backproject_points(target[i],
fx=cam[i, 2],
fy=cam[i, 3],
cx=cam[i, 0],
cy=cam[i, 1])
for j in range(masked_idx.shape[0]):
try:
images[i,
int(masked_idx[j, 0]),
int(masked_idx[j, 1]), 0] = 0
images[i,
int(masked_idx[j, 0]),
int(masked_idx[j, 1]), 1] = 255
images[i,
int(masked_idx[j, 0]),
int(masked_idx[j, 1]), 2] = 0
except:
pass
min1 = torch.min(masked_idx[:, 0])
max1 = torch.max(masked_idx[:, 0])
max2 = torch.max(masked_idx[:, 1])
min2 = torch.min(masked_idx[:, 1])
bb = BoundingBox(p1=torch.stack([min1, min2]),
p2=torch.stack([max1, max2]))
bb_img = bb.plot(images[i, :, :, :3].cpu().numpy().astype(
np.uint8))
fig.add_subplot(num, 2, i * 2 + 1)
plt.imshow(bb_img)
fig.add_subplot(num, 2, i * 2 + 2)
real = images[i, :, :, :3].cpu().numpy().astype(np.uint8)
plt.imshow(real)
if method != 'def':
a = get_img_from_fig(fig).astype(np.uint8)
plt.close()
return a
if store:
#store_ar = (img_d* 255).round().astype(np.uint8)
plt.savefig(f'{self.p_visu}/{str(epoch)}_{tag}_project_batch.png',
dpi=300)
#save_image(img_d, tag=str(epoch) + tag, p_store=self.p_visu)
if jupyter:
plt.show()
if self.writer is not None:
# you can get a high-resolution image as numpy array!!
plot_img_np = get_img_from_fig(fig)
self.writer.add_image(tag,
plot_img_np,
global_step=epoch,
dataformats='HWC')