# This is different from the TTA of official CenterNet. | |
tta_model = dict( | |
type='DetTTAModel', | |
tta_cfg=dict(nms=dict(type='nms', iou_threshold=0.5), max_per_img=100)) | |
tta_pipeline = [ | |
dict(type='LoadImageFromFile', to_float32=True, backend_args=None), | |
dict( | |
type='TestTimeAug', | |
transforms=[ | |
[ | |
# ``RandomFlip`` must be placed before ``RandomCenterCropPad``, | |
# otherwise bounding box coordinates after flipping cannot be | |
# recovered correctly. | |
dict(type='RandomFlip', prob=1.), | |
dict(type='RandomFlip', prob=0.) | |
], | |
[ | |
dict( | |
type='RandomCenterCropPad', | |
ratios=None, | |
border=None, | |
mean=[0, 0, 0], | |
std=[1, 1, 1], | |
to_rgb=True, | |
test_mode=True, | |
test_pad_mode=['logical_or', 31], | |
test_pad_add_pix=1), | |
], | |
[dict(type='LoadAnnotations', with_bbox=True)], | |
[ | |
dict( | |
type='PackDetInputs', | |
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', | |
'flip', 'flip_direction', 'border')) | |
] | |
]) | |
] | |