123456789101112131415161718192021222324252627282930313233343536373839 |
- # 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'))
- ]
- ])
- ]
|