Skip to content

Latest commit

 

History

History
119 lines (86 loc) · 4.39 KB

transforms.md

File metadata and controls

119 lines (86 loc) · 4.39 KB
Raw

Data Transforms

In this tutorial, we introduce the design of transforms pipeline in MMSegmentation.

The structure of this guide is as follows:

Design of Data pipelines

Following typical conventions, we use Dataset and DataLoader for data loading with multiple workers. Dataset returns a dict of data items corresponding the arguments of models' forward method. Since the data in semantic segmentation may not be the same size, we introduce a new DataContainer type in MMCV to help collect and distribute data of different size. See here for more details.

In 1.x version of MMSegmentation, all data transformations are inherited from BaseTransform.

The input and output types of transformations are both dict. A simple example is as follows:

>>> from mmseg.datasets.transforms import LoadAnnotations
>>> transforms = LoadAnnotations()
>>> img_path = './data/cityscapes/leftImg8bit/train/aachen/aachen_000000_000019_leftImg8bit.png.png'
>>> gt_path = './data/cityscapes/gtFine/train/aachen/aachen_000015_000019_gtFine_instanceTrainIds.png'
>>> results = dict(
>>>     img_path=img_path,
>>>     seg_map_path=gt_path,
>>>     reduce_zero_label=False,
>>>     seg_fields=[])
>>> data_dict = transforms(results)
>>> print(data_dict.keys())
dict_keys(['img_path', 'seg_map_path', 'reduce_zero_label', 'seg_fields', 'gt_seg_map'])

The data preparation pipeline and the dataset are decomposed. Usually a dataset defines how to process the annotations and a data pipeline defines all the steps to prepare a data dict. A pipeline consists of a sequence of operations. Each operation takes a dict as input and also outputs a dict for the next transform.

The operations are categorized into data loading, pre-processing, formatting and test-time augmentation.

Here is a pipeline example for PSPNet:

crop_size = (512, 1024)
train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations'),
    dict(
        type='RandomResize',
        scale=(2048, 1024),
        ratio_range=(0.5, 2.0),
        keep_ratio=True),
    dict(type='RandomCrop', crop_size=crop_size, cat_max_ratio=0.75),
    dict(type='RandomFlip', prob=0.5),
    dict(type='PhotoMetricDistortion'),
    dict(type='PackSegInputs')
]
test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='Resize', scale=(2048, 1024), keep_ratio=True),
    # add loading annotation after ``Resize`` because ground truth
    # does not need to resize data transform
    dict(type='LoadAnnotations'),
    dict(type='PackSegInputs')
]

For each operation, we list the related dict fields that are added/updated/removed. Before pipelines, the information we can directly obtain from the datasets are img_path and seg_map_path.

Data loading

LoadImageFromFile: Load an image from file.

  • add: img, img_shape, ori_shape

LoadAnnotations: Load semantic segmentation maps provided by dataset.

  • add: seg_fields, gt_seg_map

Pre-processing

RandomResize: Random resize image & segmentation map.

  • add: scale, scale_factor, keep_ratio
  • update: img, img_shape, gt_seg_map

Resize: Resize image & segmentation map.

  • add: scale, scale_factor, keep_ratio
  • update: img, gt_seg_map, img_shape

RandomCrop: Random crop image & segmentation map.

  • update: img, gt_seg_map, img_shape

RandomFlip: Flip the image & segmentation map.

  • add: flip, flip_direction
  • update: img, gt_seg_map

PhotoMetricDistortion: Apply photometric distortion to image sequentially, every transformation is applied with a probability of 0.5. The position of random contrast is in second or second to last(mode 0 or 1 below, respectively).

1. random brightness
2. random contrast (mode 0)
3. convert color from BGR to HSV
4. random saturation
5. random hue
6. convert color from HSV to BGR
7. random contrast (mode 1)
  • update: img

Formatting

PackSegInputs: Pack the inputs data for the semantic segmentation.

  • add: inputs, data_sample
  • remove: keys specified by meta_keys (merged into the metainfo of data_sample), all other keys