Module kerod.dataset.utils

None

None

View Source

from typing import Dict

import tensorflow as tf

from kerod.core.box_ops import compute_area

from kerod.core.standard_fields import BoxField

def _filter(_dict, _filter):

    keys = {BoxField.BOXES, BoxField.LABELS, BoxField.MASKS, 'is_crowd'}

    filtered_dict = {}

    for key in _dict.keys():

        if key in keys:

            filtered_dict[key] = tf.gather_nd(_dict[key], _filter)

        else:

            filtered_dict[key] = _dict[key]

    return filtered_dict

def filter_crowded_boxes(groundtruths: Dict[str, tf.Tensor]) -> Dict[str, tf.Tensor]:

    """Coco has boxes flagged as crowded which are not used during the training.

    This function will discard them.

    Arguments:

    - *groundtruths*: A dict with the following keys

        1. boxes: A tensor of shape [num_boxes, (y1, x1, y2, x2)]

        2. labels: A tensor of shape [num_boxes, ]

        3. crowd: Boolean tensor which indicates if the boxes is crowded are not.

            Crowded means that the boxes contains multiple entities which

            are to difficult to localize one by one. `True` is for crowded box.

    Returns:

    - *groundtruths*: Filtered groundtruths

    """

    ind_uncrowded_boxes = tf.where(tf.equal(groundtruths['is_crowd'], False))

    return _filter(groundtruths, ind_uncrowded_boxes)

def filter_bad_area(groundtruths: Dict[str, tf.Tensor]) -> Dict[str, tf.Tensor]:

    """Remove all the boxes that have an area less or equal to 0.

    Arguments:

    - *groundtruths*: A dict with the following keys:

        1. bbox: A tensor of shape [num_boxes, (y1, x1, y2, x2)]

        2. label: A tensor of shape [num_boxes, ]

    Returns:

    - *groundtruths*: All the groundtruths which match have not been filtered.

    """

    area = compute_area(groundtruths[BoxField.BOXES])

    filter_area = tf.where(area > 0)

    return _filter(groundtruths, filter_area)

Functions

filter_bad_area

def filter_bad_area(
    groundtruths: Dict[str, tensorflow.python.framework.ops.Tensor]
) -> Dict[str, tensorflow.python.framework.ops.Tensor]

Remove all the boxes that have an area less or equal to 0.

Arguments:

• groundtruths: A dict with the following keys:
  1. bbox: A tensor of shape [num_boxes, (y1, x1, y2, x2)]
  2. label: A tensor of shape [num_boxes, ]

Returns:

• groundtruths: All the groundtruths which match have not been filtered.

View Source

def filter_bad_area(groundtruths: Dict[str, tf.Tensor]) -> Dict[str, tf.Tensor]:

    """Remove all the boxes that have an area less or equal to 0.

    Arguments:

    - *groundtruths*: A dict with the following keys:

        1. bbox: A tensor of shape [num_boxes, (y1, x1, y2, x2)]

        2. label: A tensor of shape [num_boxes, ]

    Returns:

    - *groundtruths*: All the groundtruths which match have not been filtered.

    """

    area = compute_area(groundtruths[BoxField.BOXES])

    filter_area = tf.where(area > 0)

    return _filter(groundtruths, filter_area)

filter_crowded_boxes

def filter_crowded_boxes(
    groundtruths: Dict[str, tensorflow.python.framework.ops.Tensor]
) -> Dict[str, tensorflow.python.framework.ops.Tensor]

Coco has boxes flagged as crowded which are not used during the training.

This function will discard them.

Arguments:

• groundtruths: A dict with the following keys
  1. boxes: A tensor of shape [num_boxes, (y1, x1, y2, x2)]
  2. labels: A tensor of shape [num_boxes, ]
  3. crowd: Boolean tensor which indicates if the boxes is crowded are not. Crowded means that the boxes contains multiple entities which are to difficult to localize one by one. True is for crowded box.

Returns:

• groundtruths: Filtered groundtruths

View Source

def filter_crowded_boxes(groundtruths: Dict[str, tf.Tensor]) -> Dict[str, tf.Tensor]:

    """Coco has boxes flagged as crowded which are not used during the training.

    This function will discard them.

    Arguments:

    - *groundtruths*: A dict with the following keys

        1. boxes: A tensor of shape [num_boxes, (y1, x1, y2, x2)]

        2. labels: A tensor of shape [num_boxes, ]

        3. crowd: Boolean tensor which indicates if the boxes is crowded are not.

            Crowded means that the boxes contains multiple entities which

            are to difficult to localize one by one. `True` is for crowded box.

    Returns:

    - *groundtruths*: Filtered groundtruths

    """

    ind_uncrowded_boxes = tf.where(tf.equal(groundtruths['is_crowd'], False))

    return _filter(groundtruths, ind_uncrowded_boxes)