def eval_depth_predictions(database, save_dir, overfit=False):
# Dataloaders
if database == 'NYUD':
from data.nyud import NYUD_MT
gt_set = 'val'
db = NYUD_MT(split=gt_set, do_depth=True, overfit=overfit)
else:
raise NotImplementedError
base_name = database + '_' + 'test' + '_depth'
fname = os.path.join(save_dir, base_name + '.json')
# Eval the model
print('Evaluate the saved images (depth)')
eval_results = eval_depth(db, os.path.join(save_dir, 'depth'))
with open(fname, 'w') as f:
json.dump(eval_results, f)
# Print results
print('Results for Depth Estimation')
for x in eval_results:
spaces = ''
for j in range(0, 15 - len(x)):
spaces += ' '
print('{0:s}{1:s}{2:.4f}'.format(x, spaces, eval_results[x]))
return eval_results
def get_val_dataset(p, transforms):
""" Return the validation dataset """
db_name = p['val_db_name']
print('Preparing val loader for db: {}'.format(db_name))
if db_name == 'PASCALContext':
from data.pascal_context import PASCALContext
database = PASCALContext(split=['val'],
transform=transforms,
retname=True,
do_semseg='semseg' in p.TASKS.NAMES,
do_edge='edge' in p.TASKS.NAMES,
do_normals='normals' in p.TASKS.NAMES,
do_sal='sal' in p.TASKS.NAMES,
do_human_parts='human_parts' in p.TASKS.NAMES,
overfit=p['overfit'])
elif db_name == 'NYUD':
from data.nyud import NYUD_MT
database = NYUD_MT(split='val',
transform=transforms,
do_edge='edge' in p.TASKS.NAMES,
do_semseg='semseg' in p.TASKS.NAMES,
do_normals='normals' in p.TASKS.NAMES,
do_depth='depth' in p.TASKS.NAMES,
overfit=p['overfit'])
else:
raise NotImplemented(
"test_db_name: Choose among PASCALContext and NYUD")
return database
def eval_semseg_predictions(database, save_dir, overfit=False):
""" Evaluate the segmentation maps that are stored in the save dir """
# Dataloaders
if database == 'PASCALContext':
from data.pascal_context import PASCALContext
n_classes = 20
cat_names = VOC_CATEGORY_NAMES
has_bg = True
gt_set = 'val'
db = PASCALContext(split=gt_set,
do_edge=False,
do_human_parts=False,
do_semseg=True,
do_normals=False,
overfit=overfit)
elif database == 'NYUD':
from data.nyud import NYUD_MT
n_classes = 40
cat_names = NYU_CATEGORY_NAMES
has_bg = False
gt_set = 'val'
db = NYUD_MT(split=gt_set, do_semseg=True, overfit=overfit)
else:
raise NotImplementedError
base_name = database + '_' + 'test' + '_semseg'
fname = os.path.join(save_dir, base_name + '.json')
# Eval the model
print('Evaluate the saved images (semseg)')
eval_results = eval_semseg(db,
os.path.join(save_dir, 'semseg'),
n_classes=n_classes,
has_bg=has_bg)
with open(fname, 'w') as f:
json.dump(eval_results, f)
# Print results
class_IoU = eval_results['jaccards_all_categs']
mIoU = eval_results['mIoU']
print('\nSemantic Segmentation mIoU: {0:.4f}\n'.format(100 * mIoU))
for i in range(len(class_IoU)):
spaces = ''
for j in range(0, 15 - len(cat_names[i])):
spaces += ' '
print('{0:s}{1:s}{2:.4f}'.format(cat_names[i], spaces,
100 * class_IoU[i]))
return eval_results
def eval_normals_predictions(database, save_dir, overfit=False):
""" Evaluate the normals maps that are stored in the save dir """
# Dataloaders
if database == 'PASCALContext':
from data.pascal_context import PASCALContext
gt_set = 'val'
db = PASCALContext(split=gt_set,
do_edge=False,
do_human_parts=False,
do_semseg=False,
do_normals=True,
overfit=overfit)
elif database == 'NYUD':
from data.nyud import NYUD_MT
gt_set = 'val'
db = NYUD_MT(split=gt_set, do_normals=True, overfit=overfit)
else:
raise NotImplementedError
base_name = database + '_' + 'test' + '_normals'
fname = os.path.join(save_dir, base_name + '.json')
# Eval the model
print('Evaluate the saved images (surface normals)')
eval_results = eval_normals(db, os.path.join(save_dir, 'normals'))
with open(fname, 'w') as f:
json.dump(eval_results, f)
# Print results
print('Results for Surface Normal Estimation')
for x in eval_results:
spaces = ""
for j in range(0, 15 - len(x)):
spaces += ' '
print('{0:s}{1:s}{2:.4f}'.format(x, spaces, eval_results[x]))
return eval_results