Spaces:
Sleeping
Sleeping
File size: 15,351 Bytes
b34d1d6 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 |
# Copyright (c) OpenMMLab. All rights reserved.
import copy
from typing import Dict, List, Optional, Sequence, Union
import numpy as np
import torch
from mmengine.model import BaseDataPreprocessor
from torch import Tensor
import torch.nn.functional as F
from mmdet.structures.bbox import BaseBoxes
from mmengine.model.utils import stack_batch
from mmdet.models.utils.misc import samplelist_boxtype2tensor, unfold_wo_center
from mmdet.registry import MODELS
from mmdet.structures import TrackDataSample, TrackSampleList
from mmdet.structures.mask import BitmapMasks
from mmdet.models.data_preprocessors import DetDataPreprocessor
from mmengine.structures import PixelData
try:
import skimage
except ImportError:
skimage = None
@MODELS.register_module()
class VideoSegDataPreprocessor(DetDataPreprocessor):
"""Image pre-processor for tracking tasks.
Accepts the data sampled by the dataloader, and preprocesses
it into the format of the model input. ``TrackDataPreprocessor``
provides the tracking data pre-processing as follows:
- Collate and move data to the target device.
- Pad inputs to the maximum size of current batch with defined
``pad_value``. The padding size can be divisible by a defined
``pad_size_divisor``
- Stack inputs to inputs.
- Convert inputs from bgr to rgb if the shape of input is (1, 3, H, W).
- Normalize image with defined std and mean.
- Do batch augmentations during training.
- Record the information of ``batch_input_shape`` and ``pad_shape``.
Args:
mean (Sequence[Number], optional): The pixel mean of R, G, B
channels. Defaults to None.
std (Sequence[Number], optional): The pixel standard deviation of
R, G, B channels. Defaults to None.
pad_size_divisor (int): The size of padded image should be
divisible by ``pad_size_divisor``. Defaults to 1.
pad_value (Number): The padded pixel value. Defaults to 0.
pad_mask (bool): Whether to pad instance masks. Defaults to False.
mask_pad_value (int): The padded pixel value for instance masks.
Defaults to 0.
bgr_to_rgb (bool): whether to convert image from BGR to RGB.
Defaults to False.
rgb_to_bgr (bool): whether to convert image from RGB to RGB.
Defaults to False.
use_det_processor: (bool): whether to use DetDataPreprocessor
in training phrase. This is mainly for some tracking models
fed into one image rather than a group of image in training.
Defaults to False.
. boxtype2tensor (bool): Whether to convert the ``BaseBoxes`` type of
bboxes data to ``Tensor`` type. Defaults to True.
batch_augments (list[dict], optional): Batch-level augmentations
"""
def __init__(self,
mean: Optional[Sequence[Union[float, int]]] = None,
std: Optional[Sequence[Union[float, int]]] = None,
use_det_processor: bool = False,
**kwargs):
super().__init__(mean=mean, std=std, **kwargs)
self.use_det_processor = use_det_processor
if mean is not None and not self.use_det_processor:
# overwrite the ``register_bufffer`` in ``ImgDataPreprocessor``
# since the shape of ``mean`` and ``std`` in tracking tasks must be
# (T, C, H, W), which T is the temporal length of the video.
self.register_buffer('mean',
torch.tensor(mean).view(1, -1, 1, 1), False)
self.register_buffer('std',
torch.tensor(std).view(1, -1, 1, 1), False)
def forward(self, data: dict, training: bool = False) -> Dict:
"""Perform normalization、padding and bgr2rgb conversion based on
``TrackDataPreprocessor``.
Args:
data (dict): data sampled from dataloader.
training (bool): Whether to enable training time augmentation.
Returns:
Tuple[Dict[str, List[torch.Tensor]], OptSampleList]: Data in the
same format as the model input.
"""
if not isinstance(data['data_samples'][0], TrackDataSample):
use_det = True
else:
use_det = False
if self.use_det_processor and training or use_det:
batch_pad_shape = self._get_pad_shape(data)
else:
batch_pad_shape = self._get_track_pad_shape(data)
data = self.cast_data(data)
imgs, data_samples = data['inputs'], data['data_samples']
if self.use_det_processor and training or use_det:
assert imgs[0].dim() == 3, \
'Only support the 3 dims when use detpreprocessor in training'
if self._channel_conversion:
imgs = [_img[[2, 1, 0], ...] for _img in imgs]
# Convert to `float`
imgs = [_img.float() for _img in imgs]
if self._enable_normalize:
imgs = [(_img - self.mean.squeeze(0)) / self.std.squeeze(0) for _img in imgs]
inputs = stack_batch(imgs, self.pad_size_divisor, self.pad_value)
else:
assert imgs[0].dim() == 4, \
'Only support the 4 dims when use trackprocessor in training'
# The shape of imgs[0] is (T, C, H, W).
channel = imgs[0].size(1)
if self._channel_conversion and channel == 3:
imgs = [_img[:, [2, 1, 0], ...] for _img in imgs]
# change to `float`
imgs = [_img.float() for _img in imgs]
if self._enable_normalize:
imgs = [(_img - self.mean) / self.std for _img in imgs]
inputs = stack_track_batch(imgs, self.pad_size_divisor,
self.pad_value)
if data_samples is not None:
# NOTE the batched image size information may be useful, e.g.
# in DETR, this is needed for the construction of masks, which is
# then used for the transformer_head.
batch_input_shape = tuple(inputs.size()[-2:])
if self.use_det_processor and training or use_det:
for data_sample, pad_shape in zip(data_samples,
batch_pad_shape):
data_sample.set_metainfo({
'batch_input_shape': batch_input_shape,
'pad_shape': pad_shape
})
if self.boxtype2tensor:
samplelist_boxtype2tensor(data_samples)
if self.pad_mask and training:
self.pad_gt_masks(data_samples)
if self.pad_seg and training:
self.pad_gt_sem_seg(data_samples)
else:
for track_data_sample, pad_shapes in zip(
data_samples, batch_pad_shape):
for i in range(len(track_data_sample)):
det_data_sample = track_data_sample[i]
det_data_sample.set_metainfo({
'batch_input_shape': batch_input_shape,
'pad_shape': pad_shapes[i]
})
if self.boxtype2tensor:
tracking_samplelist_boxtype2tensor(data_samples)
if self.pad_mask and training:
self.pad_track_gt_masks(data_samples)
if self.pad_seg and training:
self.pad_track_gt_sem_seg(data_samples)
if training and self.batch_augments is not None:
for batch_aug in self.batch_augments:
if self.use_det_processor and training or use_det:
inputs, data_samples = batch_aug(inputs, data_samples)
else:
# For video segmentation, the batch augmentation are conducted
# on the batch dimension only, which means it will be run several
# times given the number of frames.
final_inputs = []
for frame_id in range(inputs.size(1)):
det_data_samples = [
track_data_sample[frame_id]
for track_data_sample in data_samples
]
aug_inputs, aug_det_samples = batch_aug(
inputs[:, frame_id], det_data_samples)
final_inputs.append(aug_inputs.unsqueeze(1))
for track_data_sample, det_sample in zip(
data_samples, aug_det_samples):
track_data_sample.video_data_samples[frame_id] = det_sample
inputs = torch.cat(final_inputs, dim=1)
# Note: inputs may contain large number of frames, so we must make
# sure that the mmeory is contiguous for stable forward
inputs = inputs.contiguous()
return dict(inputs=inputs, data_samples=data_samples)
def _get_track_pad_shape(self, data: dict) -> Dict[str, List]:
"""Get the pad_shape of each image based on data and pad_size_divisor.
Args:
data (dict): Data sampled from dataloader.
Returns:
Dict[str, List]: The shape of padding.
"""
batch_pad_shape = dict()
batch_pad_shape = []
for imgs in data['inputs']:
# The sequence images in one sample among a batch have the same
# original shape
pad_h = int(np.ceil(imgs.shape[-2] /
self.pad_size_divisor)) * self.pad_size_divisor
pad_w = int(np.ceil(imgs.shape[-1] /
self.pad_size_divisor)) * self.pad_size_divisor
pad_shapes = [(pad_h, pad_w)] * imgs.size(0)
batch_pad_shape.append(pad_shapes)
return batch_pad_shape
def pad_track_gt_masks(self,
data_samples: Sequence[TrackDataSample]) -> None:
"""Pad gt_masks to shape of batch_input_shape."""
if 'masks' in data_samples[0][0].get('gt_instances', None):
for track_data_sample in data_samples:
for i in range(len(track_data_sample)):
det_data_sample = track_data_sample[i]
masks = det_data_sample.gt_instances.masks
# TODO: whether to use BitmapMasks
assert isinstance(masks, BitmapMasks)
batch_input_shape = det_data_sample.batch_input_shape
det_data_sample.gt_instances.masks = masks.pad(
batch_input_shape, pad_val=self.mask_pad_value)
def pad_track_gt_sem_seg(self,
data_samples: Sequence[TrackDataSample]) -> None:
"""Pad gt_sem_seg to shape of batch_input_shape."""
if 'gt_sem_seg' in data_samples[0][0]:
for track_data_sample in data_samples:
for i in range(len(track_data_sample)):
det_data_sample = track_data_sample[i]
gt_sem_seg = det_data_sample.gt_sem_seg.sem_seg
h, w = gt_sem_seg.shape[-2:]
pad_h, pad_w = det_data_sample.batch_input_shape
gt_sem_seg = F.pad(
gt_sem_seg,
pad=(0, max(pad_w - w, 0), 0, max(pad_h - h, 0)),
mode='constant',
value=self.seg_pad_value)
det_data_sample.gt_sem_seg = PixelData(sem_seg=gt_sem_seg)
def stack_track_batch(tensors: List[torch.Tensor],
pad_size_divisor: int = 0,
pad_value: Union[int, float] = 0) -> torch.Tensor:
"""Stack multiple tensors to form a batch and pad the images to the max
shape use the right bottom padding mode in these images. If
``pad_size_divisor > 0``, add padding to ensure the common height and width
is divisible by ``pad_size_divisor``. The difference between this function
and ``stack_batch`` in MMEngine is that this function can process batch
sequence images with shape (N, T, C, H, W).
Args:
tensors (List[Tensor]): The input multiple tensors. each is a
TCHW 4D-tensor. T denotes the number of key/reference frames.
pad_size_divisor (int): If ``pad_size_divisor > 0``, add padding
to ensure the common height and width is divisible by
``pad_size_divisor``. This depends on the model, and many
models need a divisibility of 32. Defaults to 0
pad_value (int, float): The padding value. Defaults to 0
Returns:
Tensor: The NTCHW 5D-tensor. N denotes the batch size.
"""
assert isinstance(tensors, list), \
f'Expected input type to be list, but got {type(tensors)}'
assert len(set([tensor.ndim for tensor in tensors])) == 1, \
f'Expected the dimensions of all tensors must be the same, ' \
f'but got {[tensor.ndim for tensor in tensors]}'
assert tensors[0].ndim == 4, f'Expected tensor dimension to be 4, ' \
f'but got {tensors[0].ndim}'
assert len(set([tensor.shape[0] for tensor in tensors])) == 1, \
f'Expected the channels of all tensors must be the same, ' \
f'but got {[tensor.shape[0] for tensor in tensors]}'
tensor_sizes = [(tensor.shape[-2], tensor.shape[-1]) for tensor in tensors]
max_size = np.stack(tensor_sizes).max(0)
if pad_size_divisor > 1:
# the last two dims are H,W, both subject to divisibility requirement
max_size = (
max_size +
(pad_size_divisor - 1)) // pad_size_divisor * pad_size_divisor
padded_samples = []
for tensor in tensors:
padding_size = [
0, max_size[-1] - tensor.shape[-1], 0,
max_size[-2] - tensor.shape[-2]
]
if sum(padding_size) == 0:
padded_samples.append(tensor)
else:
padded_samples.append(F.pad(tensor, padding_size, value=pad_value))
return torch.stack(padded_samples, dim=0)
def tracking_samplelist_boxtype2tensor(batch_track_samples: TrackSampleList) -> None:
for track_data_sample in batch_track_samples:
for data_samples in track_data_sample.video_data_samples:
if 'gt_instances' in data_samples:
bboxes = data_samples.gt_instances.get('bboxes', None)
if isinstance(bboxes, BaseBoxes):
data_samples.gt_instances.bboxes = bboxes.tensor
if 'pred_instances' in data_samples:
bboxes = data_samples.pred_instances.get('bboxes', None)
if isinstance(bboxes, BaseBoxes):
data_samples.pred_instances.bboxes = bboxes.tensor
if 'ignored_instances' in data_samples:
bboxes = data_samples.ignored_instances.get('bboxes', None)
if isinstance(bboxes, BaseBoxes):
data_samples.ignored_instances.bboxes = bboxes.tensor
|