Add transform to change dataset image resolution

[kankanzheli]

I annotated the dataset using cvat. But I found that the resolution of my image was too high. Is there any way to reduce the resolution of the marked image?

[nmanovic]

@kankanzheli , you can try Datumaro (https://github.com/openvinotoolkit/datumaro) for the task.

@zhiltsov-max , could you please advise?

[zhiltsov-max]

Hi, could you say the format name?

[kankanzheli]

coco

[kankanzheli]

I am working in coco format. And I did the labeled work by cvat.

[zhiltsov-max]

Ok, try the following script:

from datumaro.components.dataset import Dataset
from datumaro.components.annotation import Bbox, Label, Mask, Points, PolyLine, Polygon, RleMask
from datumaro.components.extractor import DatasetItem, ItemTransform
from datumaro.util.mask_tools import mask_to_rle

import pycocotools.mask as mask_util

import cv2
import numpy as np

class ScaleTransform(ItemTransform):
    def __init__(self, extractor, scale=1):
        super().__init__(extractor)

        self._scale = scale

    def transform_item(self, item):
        if item.has_image and item.image.has_data:
            method = cv2.INTER_LINEAR if self._scale < 1 else cv2.INTER_LANCZOS4
            image = item.image.data / 255.0
            rescaled_image = cv2.resize(image, (0, 0),
                fx=self._scale, fy=self._scale, interpolation=method)

        rescaled_annotations = []
        for ann in item.annotations:
            if isinstance(ann, Bbox):
                rescaled_annotations.append(ann.wrap(
                    x=ann.x * self._scale,
                    y=ann.y * self._scale,
                    w=ann.w * self._scale,
                    h=ann.h * self._scale,
                ))
            elif isinstance(ann, (Polygon, Points, PolyLine)):
                rescaled_annotations.append(ann.wrap(
                    points=[p * self._scale for p in ann.points]
                ))
            elif isinstance(ann, Mask):
                # Can only use NEAREST for masks
                rescaled_mask = cv2.resize(ann.image, (0, 0),
                    fx=self._scale, fy=self._scale,
                    interpolation=cv2.INTER_NEAREST)

                if isinstance(ann, RleMask):
                    rle = mask_to_rle(rescaled_mask)
                    rescaled_annotations.append(ann.wrap(
                        rle=mask_util.frPyObjects(rle, *rle['size'])))
                else:
                    rescaled_annotations.append(ann.wrap(image=rescaled_mask))
            else:
                rescaled_annotations.append(ann)

        return self.wrap_item(item,
            image=rescaled_image, annotations=rescaled_annotations)

dataset = Dataset.import_from('path/', 'coco')
dataset.transform(ScaleTransform, scale=0.5) # you can specify the desired scale here
dataset.export('path2/', 'coco', save_images=True)

@kirill-sizov, please integrate this transform into Datumaro, it looks quite useful.

[kankanzheli]

How to get the rescaled image at the same time?

[kankanzheli]

I use the annotation result (JSON file) obtained by this method on the image whose resolution I modified by other methods. At this time, the annotation result is misaligne

d.

[kankanzheli]

I only modify these four parameters during use:

rescaled_ image = cv2.resize(image, (1280, 720)

rescaled_mask = cv2.resize(ann.image, (1280, 720),

dataset = Dataset.import_from('/home/zhaowentao/coco/', 'coco')

dataset.export('/home/zhaowentao/1/', 'coco', save_images=True)

[zhiltsov-max]

If you have fixed resulting image size, try this one instead:

from datumaro.components.dataset import Dataset
from datumaro.components.annotation import Bbox, Label, Mask, Points, PolyLine, Polygon, RleMask
from datumaro.components.extractor import DatasetItem, ItemTransform
from datumaro.util.mask_tools import mask_to_rle
from datumaro.util import take_by

import pycocotools.mask as mask_util

import cv2
import numpy as np

class ScaleTransform(ItemTransform):
    def __init__(self, extractor, size):
        super().__init__(extractor)

        self._size = size

    def transform_item(self, item):
        assert item.has_image
        h, w = item.image.size
        xscale = self._size[0] / w
        yscale = self._size[1] / h

        if item.image.has_data:
            method = cv2.INTER_LINEAR if (xscale * yscale) < 1 else cv2.INTER_LANCZOS4
            image = item.image.data / 255.0
            rescaled_image = cv2.resize(image, (0, 0),
                fx=xscale, fy=yscale, interpolation=method)

        rescaled_annotations = []
        for ann in item.annotations:
            if isinstance(ann, Bbox):
                rescaled_annotations.append(ann.wrap(
                    x=ann.x * xscale,
                    y=ann.y * yscale,
                    w=ann.w * xscale,
                    h=ann.h * yscale,
                ))
            elif isinstance(ann, (Polygon, Points, PolyLine)):
                rescaled_annotations.append(ann.wrap(
                    points=[p
                        for t in ((x * xscale, y * yscale) for x, y in take_by(ann.points, 2))
                        for p in t
                    ]
                ))
            elif isinstance(ann, Mask):
                # Can only use NEAREST for masks
                rescaled_mask = cv2.resize(ann.image, (0, 0),
                    fx=xscale, fy=yscale,
                    interpolation=cv2.INTER_NEAREST)

                if isinstance(ann, RleMask):
                    rle = mask_to_rle(rescaled_mask)
                    rescaled_annotations.append(ann.wrap(
                        rle=mask_util.frPyObjects(rle, *rle['size'])))
                else:
                    rescaled_annotations.append(ann.wrap(image=rescaled_mask))
            else:
                rescaled_annotations.append(ann)

        return self.wrap_item(item,
            image=rescaled_image,
            annotations=rescaled_annotations)

dataset = Dataset.import_from('path/', 'coco')
dataset.transform(ScaleTransform, size=(1280, 720))
dataset.export('path2/', 'coco', save_images=True)

[kankanzheli]

error is happened:
File "Change_resolution.py", line 41, in transform_item

points=[p for p in (x * xscale, y * yscale)

NameError: name 'x' is not defined

[zhiltsov-max]

I've updated the answer above.

[kankanzheli]

It is worked. Thanks！