def visualize_results(self, batch, out): background_image = batch["scene_img"][0]["scaled_image"].copy() inp = batch["in_xy"] gt = batch["gt_xy"] pred = out["out_xy"] pred = pred.view(pred.size(0), self.hparams.best_k_val, -1, pred.size(-1)) y = out["y_map"] y_softmax = out["y_softmax"] image = visualize_traj_probabilities( input_trajectory=inp.cpu()[:, 0], gt_trajectory=None, prediction_trajectories=pred.cpu()[:,:, 0], background_image=background_image, img_scaling=self.val_dset.img_scaling, scaling_global=self.val_dset.scaling_global, grid_size=20, y_softmax=y_softmax, y=y, global_patch=re_im(batch["global_patch"][0]).cpu().numpy(), probability_mask=batch["prob_mask"][0][0].cpu().numpy(), grid_size_in_global=self.val_dset.grid_size_in_global ) self.logger.experiment.add_image(f'Trajectories', image, self.current_epoch)
def visualize_results(self, batch, out):
y = out["y_map"]
y_softmax = out["y_softmax"]
image = visualize_probabilities(
y_softmax=y_softmax,
y=y,
global_patch=re_im(batch["global_patch"][0]).cpu().numpy(),
probability_mask=batch["prob_mask"][0][0].cpu().numpy(),
grid_size_in_global=self.val_dset.grid_size_in_global)
self.logger.experiment.add_image(f'Map', image, self.current_epoch)
def plot(self,
index,
modes=["in", "gt"],
image_type="scaled",
final_mask=False):
out = self.get_scene(index)
if "seq_start_end" in out:
(index, end) = out["seq_start_end"]
image = out["scene_img"][0]
if image_type == "orig":
img_label = "img"
img = image[img_label]
scale = 1
elif image_type == "scaled":
img_label = "scaled_image"
img = image[img_label]
if self.format == "meter":
scale = 1. / self.img_scaling
else:
scale = 1
elif image_type == "global":
img_label = "global_image"
img = image[img_label]
scale = 1. / self.scaling_global
elif image_type == "local":
img_label = "local_image"
img = image[img_label]
scale = 1. / self.scaling_local
elif image_type == "patch":
img = re_im(out["global_patch"][0, :3].permute(1, 2, 0))
scale = 1. / self.scaling_global
else:
assert False, "'{}' not valid <image_type>".format(image_type)
center = out["in_xy"][-1, 0] * scale
fig = plt.figure()
ax = fig.add_subplot(111)
ax.imshow(img)
if final_mask:
rel_scaling = (2 * self.grid_size_in_global +
1) / (2 * self.grid_size_out_global + 1)
mask = out["prob_mask"][0, 0]
final_pos_out = torch.cat(torch.where(mask == 1)).float()
final_pos = torch.zeros_like(final_pos_out)
final_pos[0] = final_pos_out[1]
final_pos[1] = final_pos_out[0]
center_pixel_global = out["in_xy"][-1, 0] * scale
if not image_type == "patch":
final_pos_pixel_centered = (
final_pos - self.grid_size_out_global
) * self.scaling_global * scale * rel_scaling
final_pos_pixel = final_pos_pixel_centered + center_pixel_global
elif image_type == "patch":
final_pos_pixel = final_pos * rel_scaling
mask = mask.numpy()
mask = (mask * 255).astype(np.uint8)
mask = Image.fromarray(mask, mode="L")
mask = mask.resize((int(2 * self.grid_size_in_global + 1),
int(2 * self.grid_size_in_global + 1)))
ax.imshow(mask, alpha=0.5)
plt.plot(final_pos_pixel[0], final_pos_pixel[1], "x")
error = abs(final_pos_pixel / scale - out["gt_xy"][-1])
error_bound = rel_scaling * self.scaling_global
print("Error real and approximation: {} Threshold: {} ".format(
error, error_bound))
# assert (error < error_bound).all(), "Error above error bound"
for m in modes:
if m == "gt":
marker = '-'
else:
marker = '-'
traj = out["{}_xy".format(m)][:, 0] * scale
traj = traj.cpu().numpy()
if image_type == "patch":
traj = traj + (self.grid_size_in_global) - center.cpu().numpy()
ax.plot((traj[:, 0]).astype(int), (traj[:, 1]).astype(int),
linestyle=marker,
linewidth=int(3))
plt.show()
def trim_axs(axs, N):
"""little helper to massage the axs list to have correct length..."""
axs = axs.flat
for ax in axs[N:]:
ax.remove()
return axs[:N]