def layout_analysis_output(args):
ground_truth_folder = args.ground_truth_folder
network_output_folder = args.network_output_folder
output_folder = os.path.join(args.output_folder,
'layout_analysis_evaluation')
gt_img_paths = get_img_paths(ground_truth_folder)
segm_img_paths = get_img_paths(network_output_folder)
if not output_folder:
output_folder = os.path.join(os.path.dirname(segm_img_paths),
'layout_analysis_evaluation')
for gt, segm in zip(gt_img_paths, segm_img_paths):
gt_image = np.array(pil_loader(gt))
segm_image = np.array(pil_loader(segm))
generate_layout_analysis_output(output_folder, gt_image, segm_image,
os.path.basename(gt))
# create a legend and save it
make_colour_legend_image(
os.path.join(output_folder, "layout_analysis_eval_legend"),
CLASS_COLOUR_ENCODINGS)
def generate_layout_analysis_output(output_folder,
gt_image,
segm_image,
output_name,
legend=False):
"""
This function generates and saves an output like the one generated from the Output image as described in https://github.com/DIVA-DIA/DIVA_Layout_Analysis_Evaluator
Parameters
----------
output_folder: string
folder where the output is saved to
gt_image: numpy array
ground truth image in RGB
segm_image: numpy array
segmentation output in RGB
output_name: string
name of the output image
legend: Boolaen
set to true if legend should also be generated
Returns
-------
None (saves output)
"""
dest_filename = os.path.join(output_folder, output_name)
if not os.path.exists(os.path.dirname(dest_filename)):
os.makedirs(os.path.dirname(dest_filename))
img_la = np.zeros(segm_image.shape)
# Extract just blue channel
out_blue = segm_image[:, :, 2] #[2140:2145, 1570:1575]
gt_blue = gt_image[:, :, 2] #[2140:2145, 1570:1575]
# subtract the boundary pixel from the gt
boundary_pixel = gt_image[:, :, 0].astype(np.uint8) == 128
gt_blue[boundary_pixel] = 1
# get the colour masks
masks = {
c: _get_mask(c, out_blue, gt_blue)
for c in CLASS_COLOUR_ENCODINGS.keys()
}
# colour the pixels according to the masks
for c, mask in masks.items():
img_la[mask] = CLASS_COLOUR_ENCODINGS[c]
# correct for boundary pixels
boundary_masks = _get_boundary_masks(out_blue, gt_blue, boundary_pixel)
# colour the pixels according to the masks
for c, mask in boundary_masks.items():
img_la[mask] = CLASS_COLOUR_ENCODINGS[c]
# save the output
result = Image.fromarray(img_la.astype(np.uint8))
result.save(dest_filename)
if legend:
make_colour_legend_image(
os.path.join(os.path.dirname(dest_filename),
"layout_analysis_eval_legend"),
CLASS_COLOUR_ENCODINGS)
def _save_output_evaluation(class_encodings, output_encoded, gt_image, tag, multi_run=None):
"""Utility function to save image in the output folder and also log it to Tensorboard.
Parameters
----------
class_encodings : List
Contains the range of encoded classes
tag : str
Name of the image.
output_encoded : ndarray [W x H x C] in RGB
Image to be saved
gt_image : ndarray [W x H x C] in RGB
Image to be saved
multi_run : int
Epoch/Mini-batch counter.
Returns
-------
None
"""
# ##################################################################################################################
# 1. Create true output
# Get output folder using the FileHandler from the logger.
# (Assumes the file handler is the last one)
output_folder = os.path.dirname(logging.getLogger().handlers[-1].baseFilename)
dest_filename = os.path.join(output_folder, 'images', "output", tag if multi_run is None else tag + '_{}'.format(multi_run))
if not os.path.exists(os.path.dirname(dest_filename)):
os.makedirs(os.path.dirname(dest_filename))
# Save the output
Image.fromarray(output_encoded.astype(np.uint8)).save(dest_filename)
# ##################################################################################################################
# 2. Make a more human readable output -> one colour per class
tag_col = "coloured/" + tag
dest_filename = os.path.join(output_folder, 'images', tag_col if multi_run is None else tag_col + '_{}'.format(multi_run))
if not os.path.exists(os.path.dirname(dest_filename)):
os.makedirs(os.path.dirname(dest_filename))
img = np.copy(output_encoded)
blue = output_encoded[:, :, 2] # Extract just blue channel
# Colours are in RGB
cmap = matplotlib.cm.get_cmap('Spectral')
colors = [cmap(i / len(class_encodings), bytes=True)[:3] for i in range(len(class_encodings))]
# Get the mask for each colour
masks = {color: (blue == i) > 0 for color, i in zip(colors, class_encodings)}
# Color the image with relative colors
for color, mask in masks.items():
img[mask] = color
# Make and save the class color encoding
color_encoding = {str(i): color for color, i in zip(colors, class_encodings)}
make_colour_legend_image(os.path.join(os.path.dirname(dest_filename), "output_visualizations_colour_legend.png"),
color_encoding)
# Write image to output folder
Image.fromarray(img.astype(np.uint8)).save(dest_filename)
# ##################################################################################################################
# 3. Layout analysis evaluation
# Output image as described in https://github.com/DIVA-DIA/DIVA_Layout_Analysis_Evaluator
tag_la = "layout_analysis_evaluation/" + tag
dest_filename = os.path.join(output_folder, 'images', tag_la if multi_run is None else tag_la + '_{}'.format(multi_run))
if not os.path.exists(os.path.dirname(dest_filename)):
os.makedirs(os.path.dirname(dest_filename))
generate_layout_analysis_output(os.path.join(output_folder, 'images'), gt_image, output_encoded, dest_filename, legend=True)