def make_detmot_transforms(image_set, args=None):
normalize = T.MotCompose([
T.MotToTensor(),
T.MotNormalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
scales = [480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800]
if image_set == 'train':
color_transforms = []
scale_transforms = [
T.MotRandomHorizontalFlip(),
T.MotRandomResize(scales, max_size=1333),
normalize,
]
return T.MotCompose(color_transforms + scale_transforms)
if image_set == 'val':
return T.MotCompose([
T.MotRandomResize([800], max_size=1333),
normalize,
])
raise ValueError(f'unknown {image_set}')
def make_transforms_for_crowdhuman(image_set, args=None):
normalize = T.MotCompose([
T.MotToTensor(),
T.MotNormalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
scales = [608, 640, 672, 704, 736, 768, 800, 832, 864, 896, 928, 960, 992]
if image_set == 'train':
return T.MotCompose([
T.MotRandomHorizontalFlip(),
T.FixedMotRandomShift(bs=1),
T.MotRandomSelect(
T.MotRandomResize(scales, max_size=1536),
T.MotCompose([
T.MotRandomResize([400, 500, 600]),
T.FixedMotRandomCrop(384, 600),
T.MotRandomResize(scales, max_size=1536),
])
),
normalize,
])
if image_set == 'val':
return T.MotCompose([
T.MotRandomResize([800], max_size=1333),
normalize,
])
raise ValueError(f'unknown {image_set}')
def make_detmot_transforms(image_set, args=None):
normalize = T.MotCompose([
T.MotToTensor(),
T.MotNormalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
scales = [608, 640, 672, 704, 736, 768, 800, 832, 864, 896, 928, 960, 992]
if image_set == 'train':
color_transforms = []
if args.cj:
print('Training with RandomColorJitter.')
color_transforms.append(
T.MoTColorJitter(brightness=0.5,
contrast=0.5,
saturation=0.5,
hue=0))
if not args.crop:
scale_transforms = [
T.MotRandomHorizontalFlip(),
T.FixedMotRandomShift(bs=1),
T.MotRandomResize(scales, max_size=1536),
normalize,
]
else:
print('Training with RandomCrop.')
scale_transforms = [
T.MotRandomHorizontalFlip(),
T.FixedMotRandomShift(bs=1),
T.MotRandomSelect(
T.MotRandomResize(scales, max_size=1536),
T.MotCompose([
T.MotRandomResize([400, 500, 600]),
T.FixedMotRandomCrop(384, 600),
T.MotRandomResize(scales, max_size=1536),
])),
normalize,
]
return T.MotCompose(color_transforms + scale_transforms)
if image_set == 'val':
return T.MotCompose([
T.MotRandomResize([800], max_size=1536),
normalize,
])
raise ValueError(f'unknown {image_set}')