Skip to content

lacss.data

Lacss Data Pipeline API

TF Dataset

dataset_from_coco_annotations(annotation_file, image_path, image_shape=[None, None, 3], mask_shape=[48, 48])

Obtaining a tensowflow dataset from coco annotations. See coco_generator_full()

Parameters:

Name Type Description Default
annotation_file str

Path to coco annotation files

 required
image_path str

Path to image directory

 required
image_shape tuple

The expect image shapes. Use None to represent variable dimensions.

 [None, None, 3]
mask_shape tuple

If supplied, all the instance segmentations will be croped and resized to the specifed size. Otherwise, the segmentations are uncropped (in original image size)

 [48, 48]

Returns:

Type Description
Dataset

A tensorflow dataset.

dataset_from_simple_annotations(annotation_file, image_path, image_shape=[None, None, 3])

Obtaining a tensowflow dataset from simple annotatiion. See simple_generator()

Parameters:

Name Type Description Default
annotation_file str

Path to the json format annotation file.

 required
image_path str

Path to the image directory.

 required
image_shape

The expect image shapes. Use None to represent variable dimensions.

 [None, None, 3]

Returns:

Type Description
Dataset

A tensorflow dataset object

dataset_from_img_mask_pairs(imgfiles, maskfiles, *, image_shape=[None, None, 3], generate_masks=False, mask_shape=[48, 48])

Obtaining a tensowflow dataset from image/label pairs. See img_mask_pair_generator()

Parameters:

Name Type Description Default
imgfiles Sequence[str | Path]

List of file pathes to input image file.

 required
maskfiles Sequence[str | Path]

List of file pathes to label image file.

 required

Other Parameters:

Name Type Description
image_shape Sequence[int | None]

The expect image shapes. Use None to represent variable dimensions.
generate_masks bool

Whether to convert label images to indiviudal instance masks. This should be True if your data augmentation pipeline include rescaling ops or cropping ops., becasue these ops does not recompute the label image.
mask_shape tuple[int, int]

Only when generate_mask=True. The resolution of the generated masks.

Returns:

Type Description
Dataset

A tensorflow dataset object

Data augmentation

The augmentation functions for TF Dataset inputs.

crop_to_roi(inputs, *, roi, area_ratio_threshold=1.0, p=1.0)

Crop image to bounding-box ROI

Parameters:

Name Type Description Default
inputs dict

data from TF dataset. Affected elements are:

  • image
  • image_mask
  • centroids
  • bboxes
  • masks
 required

Other Parameters:

Name Type Description
roi tuple[int, int, int, int]

Rectangle roi in yxyx format
area_ratio_threshold float

remove instances if the bbox's relative remaining area is below this threshold
p float

probability of applying transformation

flip_left_right(inputs, *, p=1.0)

Flip image left-right

Parameters:

Name Type Description Default
inputs dict

data from TF dataset. Affected elements are:

  • image
  • image_mask
  • centroids
  • bboxes
 required

Other Parameters:

Name Type Description
p float

probability of applying transformation

flip_up_down(inputs, *, p=1.0)

Flip image up-down

Parameters:

Name Type Description Default
inputs dict

data from TF dataset. Affected elements are:

  • image
  • image_mask
  • centroids
  • bboxes
 required

Other Parameters:

Name Type Description
p float

probability of applying transformation

pad(inputs, *, paddings, constant_values=0, p=1.0)

Pad image and labels.

Parameters:

Name Type Description Default
inputs dict

data from TF dataset. Affected elements are:

  • image
  • image_mask
  • centroids
  • bboxes
 required

Other Parameters:

Name Type Description
paddings int | tuple[int, int]

either a tuple or a single value. If latter, use the same padding for both x and y axis
constant_values float

the value to fill the padded area
p float

probability of applying transformation

pad_to_size(inputs, *, target_size, constant_values=0, p=1.0)

Pad image and labels to a target size. Padding is applied so that the orginal scene is centered in the output.

Parameters:

Name Type Description Default
inputs dict

data from TF dataset. Affected elements are:

  • image
  • image_mask
  • centroids
  • bboxes
 required

Other Parameters:

Name Type Description
target_size tuple[int, int]

target image size
constant_values float

the value to fill the padded area
p float

probability of applying transformation

random_crop(inputs, *, target_size, area_ratio_threshold=1.0, p=1.0)

Random crop to a set target size

Parameters:

Name Type Description Default
inputs dict

data from TF dataset. Affected elements are:

  • image
  • image_mask
  • centroids
  • bboxes
  • masks
 required

Other Parameters:

Name Type Description
target_size tuple[int, int]

the target size
area_ratio_threshold float

remove instances if the bbox's relative remaining area is below this threshold
p float

probability of applying transformation

random_crop_or_pad(inputs, *, target_size, constant_values=0, area_ratio_threshold=1.0, p=1.0)

Random crop or pad image to specified target_size.

Parameters:

Name Type Description Default
inputs dict

data from TF dataset. Affected elements are:

  • image
  • image_mask
  • centroids
  • bboxes
  • masks
 required

Other Parameters:

Name Type Description
target_size tuple[int, int]

target size
constant_values float

the value to fill the padded area
area_ratio_threshold float

remove instances if the bbox's relative remaining area is below this threshold
p float

probability of applying transformation

random_resize(inputs, *, scaling, keep_aspect_ratio=False, p=1.0)

Resize image by a random amount

Parameters:

Name Type Description Default
inputs dict

data from TF dataset. Affected elements are:

  • image
  • image_mask
  • centroids
 required

Other Parameters:

Name Type Description
scaling float | tuple[float, float]

range of scale, e.g, [0.8, 1.5].
keep_aspect_ratio bool

Whether to scale x/y the same amount.
p float

probability of applying transformation

resize(inputs, *, target_size, p=1.0)

Resize image and labels

Parameters:

Name Type Description Default
inputs dict

data from TF dataset. Affected elements are:

  • image
  • image_mask
  • centroids
  • bboxes
 required

Other Parameters:

Name Type Description
target_size tuple[int, int]

target size
p float

probability of applying transformation

Python generator

coco_generator_full(annotation_file, image_path, mask_shape=None)

A generator function to produce coco-annotated data

Parameters:

Name Type Description Default
annotation_file str

Path to coco annotation files

 required
image_path str

Path to image directory

 required
mask_shape Optional[tuple[int, int]]

If supplied, all the instance segmentations will be croped and resized to the specifed size. Otherwise, the segmentations are uncropped (in original image size)

 None

Yields:

Type Description
dict

A data dictionary with thse keys:

  • id: data id
  • filename: image filename
  • image: an array [H, W, C]
  • masks: segmentation masks. [N, H, W] or [N,] + mask_shape
  • centroids: yx format.
  • bboxes: y0x0y1x1 format.
  • label: an array [H, W] representing pixel labels of all instances.

simple_generator(annotation_file, image_path)

A simple generator function to produce image data labeled with points and image-level segmentaion.

Parameters:

Name Type Description Default
annotation_file str

Path to the json format annotation file.

 required
image_path str

Path to the image directory.

 required

Yields:

Type Description
dict

A data dictionary with thse keys:

  • img_id: data id
  • image: an array [H, W, C]
  • centroids: yx format.
  • image_mask: segmentation masks for the image. [H, W]

img_mask_pair_generator(imgfiles, maskfiles)

A generator function to produce image data labeled with segmentation labels. In this case, one has paired input images and label images as files on disk.

Parameters:

Name Type Description Default
imgfiles Sequence[str | Path]

List of file pathes to input image file.

 required
maskfiles Sequence[str | Path]

List of file pathes to label image file.

 required

Yields:

Type Description
dict

A data dictionary with thse keys:

  • img_id: data id
  • image: an array [H, W, C]
  • centroids: yx format.
  • bboxes: y0x0y1x1 format.
  • label: an array [H, W] representing pixel labels of all instances.