def benchmark_segmentation(pred_dir, gt_dir, num_classes, string_replace):
"""Benchmark segmentaion on mean Intersection over Union (mIoU).
"""
print('Benchmarking semantic segmentation.')
assert(os.path.isdir(pred_dir))
assert(os.path.isdir(gt_dir))
tp_fn = np.zeros(num_classes, dtype=np.float64)
tp_fp = np.zeros(num_classes, dtype=np.float64)
tp = np.zeros(num_classes, dtype=np.float64)
for dirpath, dirnames, filenames in os.walk(pred_dir):
for filename in filenames:
predname = os.path.join(dirpath, filename)
gtname = predname.replace(pred_dir, gt_dir)
if string_replace != '':
stra, strb = string_replace.split(',')
gtname = gtname.replace(stra, strb)
pred = np.asarray(
Image.open(predname).convert(mode='L'),
dtype=np.uint8
)
gt = np.asarray(
Image.open(gtname).convert(mode='L'),
dtype=np.uint8
)
_tp_fn, _tp_fp, _tp = iou_stats(
pred,
gt,
num_classes=num_classes,
background=0
)
tp_fn += _tp_fn
tp_fp += _tp_fp
tp += _tp
iou = tp / (tp_fn + tp_fp - tp + 1e-12) * 100.0
class_names = ['beam', 'board', 'bookcase', 'ceiling', 'chair', 'clutter',
'column', 'door', 'floor', 'sofa', 'table', 'wall', 'window']
for i in range(num_classes):
print('class {:10s}: {:02d}, acc: {:4.4f}%'.format(
class_names[i], i, iou[i])
)
mean_iou = iou.sum() / num_classes
print('mean IOU: {:4.4f}%'.format(mean_iou))
mean_pixel_acc = tp.sum() / (tp_fp.sum() + 1e-12)
print('mean Pixel Acc: {:4.4f}%'.format(mean_pixel_acc))
def calcu_voc_mIou(pred_dir, gt_dir):
assert os.path.isdir(pred_dir)
assert os.path.isdir(gt_dir)
print('......')
n_class = 21
tp_fn = np.zeros(n_class, dtype=np.float64)
tp_fp = np.zeros(n_class, dtype=np.float64)
tp = np.zeros(n_class, dtype=np.float64)
for parent, dirs, files in os.walk(pred_dir):
for file in files:
pred_img_file = os.path.join(parent, file)
gt_img_file = pred_img_file.replace(pred_dir, gt_dir)
# if args.string_replace != '':
# stra, strb = args.string_replace.split(',')
# gtname = gtname.replace(stra, strb)
pred = np.asarray(Image.open(pred_img_file).convert(mode='L'),
dtype=np.uint8)
gt = np.asarray(Image.open(gt_img_file).convert(mode='P'),
dtype=np.uint8)
_tp_fn, _tp_fp, _tp = iou_stats(pred,
gt,
num_classes=n_class,
background=0)
tp_fn += _tp_fn
tp_fp += _tp_fp
tp += _tp
iou = tp / (tp_fn + tp_fp - tp + 1e-12) * 100.0
class_names = [
'Background', 'Aero', 'Bike', 'Bird', 'Boat', 'Bottle', 'Bus', 'Car',
'Cat', 'Chair', 'Cow', 'Table', 'Dog', 'Horse', 'MBike', 'Person',
'Plant', 'Sheep', 'Sofa', 'Train', 'TV'
]
for i in range(n_class):
print('class {:10s}: {:02d}, acc: {:4.4f}%'.format(
class_names[i], i, iou[i]))
mean_iou = iou.sum() / n_class
print('mean IOU: {:4.4f}%'.format(mean_iou))
mean_pixel_acc = tp.sum() / (tp_fp.sum() + 1e-12)
print('mean Pixel Acc: {:4.4f}%'.format(mean_pixel_acc))
instname = predname.replace(args.pred_dir, args.inst_dir)
if args.string_replace != '':
stra, strb = args.string_replace.split(',')
gtname = gtname.replace(stra, strb)
instname = instname.replace(stra, strb)
pred = np.asarray(Image.open(predname).convert(mode='L'),
dtype=np.uint8)
gt = np.asarray(Image.open(gtname).convert(mode='L'), dtype=np.uint8)
inst = np.asarray(Image.open(instname).convert(mode='P'),
dtype=np.uint8)
# Compute true-positive, false-positive
# and false-negative
_tp_fn, _tp_fp, _tp = iou_stats(pred,
gt,
num_classes=args.num_classes,
background=0)
# Compute num. of instances per class
inst_inds = np.unique(inst)
ninst_ = np.zeros(args.num_classes, dtype=np.float64)
for i in range(inst_inds.size):
if i < 255:
inst_ind = inst_inds[i]
inst_mask = inst == inst_ind
seg_mask = gt[inst_mask]
npixel, _ = np.histogram(
seg_mask,
bins=args.num_classes,
range=(0, args.num_classes -
1)) # num. pixel per class of this inst.