mmaction/datasets/transforms/processing.py

# Copyright (c) OpenMMLab. All rights reserved.
import random
import warnings
from numbers import Number
from typing import Sequence

import cv2
import mmcv
import mmengine
import numpy as np
from mmcv.transforms import BaseTransform
from mmcv.transforms.utils import cache_randomness
from torch.nn.modules.utils import _pair

from mmaction.registry import TRANSFORMS


def _combine_quadruple(a, b):
    return a[0] + a[2] * b[0], a[1] + a[3] * b[1], a[2] * b[2], a[3] * b[3]


def _flip_quadruple(a):
    return 1 - a[0] - a[2], a[1], a[2], a[3]


def _init_lazy_if_proper(results, lazy):
    """Initialize lazy operation properly.

    Make sure that a lazy operation is properly initialized,
    and avoid a non-lazy operation accidentally getting mixed in.

    Required keys in results are "imgs" if "img_shape" not in results,
    otherwise, Required keys in results are "img_shape", add or modified keys
    are "img_shape", "lazy".
    Add or modified keys in "lazy" are "original_shape", "crop_bbox", "flip",
    "flip_direction", "interpolation".

    Args:
        results (dict): A dict stores data pipeline result.
        lazy (bool): Determine whether to apply lazy operation. Default: False.
    """

    if 'img_shape' not in results:
        results['img_shape'] = results['imgs'][0].shape[:2]
    if lazy:
        if 'lazy' not in results:
            img_h, img_w = results['img_shape']
            lazyop = dict()
            lazyop['original_shape'] = results['img_shape']
            lazyop['crop_bbox'] = np.array([0, 0, img_w, img_h],
                                           dtype=np.float32)
            lazyop['flip'] = False
            lazyop['flip_direction'] = None
            lazyop['interpolation'] = None
            results['lazy'] = lazyop
    else:
        assert 'lazy' not in results, 'Use Fuse after lazy operations'


@TRANSFORMS.register_module()
class Fuse(BaseTransform):
    """Fuse lazy operations.

    Fusion order:
        crop -> resize -> flip

    Required keys are "imgs", "img_shape" and "lazy", added or modified keys
    are "imgs", "lazy".
    Required keys in "lazy" are "crop_bbox", "interpolation", "flip_direction".
    """

    def transform(self, results):
        """Fuse lazy operations.

        Args:
            results (dict): The resulting dict to be modified and passed
                to the next transform in pipeline.
        """
        if 'lazy' not in results:
            raise ValueError('No lazy operation detected')
        lazyop = results['lazy']
        imgs = results['imgs']

        # crop
        left, top, right, bottom = lazyop['crop_bbox'].round().astype(int)
        imgs = [img[top:bottom, left:right] for img in imgs]

        # resize
        img_h, img_w = results['img_shape']
        if lazyop['interpolation'] is None:
            interpolation = 'bilinear'
        else:
            interpolation = lazyop['interpolation']
        imgs = [
            mmcv.imresize(img, (img_w, img_h), interpolation=interpolation)
            for img in imgs
        ]

        # flip
        if lazyop['flip']:
            for img in imgs:
                mmcv.imflip_(img, lazyop['flip_direction'])

        results['imgs'] = imgs
        del results['lazy']

        return results


@TRANSFORMS.register_module()
class RandomCrop(BaseTransform):
    """Vanilla square random crop that specifics the output size.

    Required keys in results are "img_shape", "keypoint" (optional), "imgs"
    (optional), added or modified keys are "keypoint", "imgs", "lazy"; Required
    keys in "lazy" are "flip", "crop_bbox", added or modified key is
    "crop_bbox".

    Args:
        size (int): The output size of the images.
        lazy (bool): Determine whether to apply lazy operation. Default: False.
    """

    def __init__(self, size, lazy=False):
        if not isinstance(size, int):
            raise TypeError(f'Size must be an int, but got {type(size)}')
        self.size = size
        self.lazy = lazy

    @staticmethod
    def _crop_kps(kps, crop_bbox):
        """Static method for cropping keypoint."""
        return kps - crop_bbox[:2]

    @staticmethod
    def _crop_imgs(imgs, crop_bbox):
        """Static method for cropping images."""
        x1, y1, x2, y2 = crop_bbox
        return [img[y1:y2, x1:x2] for img in imgs]

    @staticmethod
    def _box_crop(box, crop_bbox):
        """Crop the bounding boxes according to the crop_bbox.

        Args:
            box (np.ndarray): The bounding boxes.
            crop_bbox(np.ndarray): The bbox used to crop the original image.
        """

        x1, y1, x2, y2 = crop_bbox
        img_w, img_h = x2 - x1, y2 - y1

        box_ = box.copy()
        box_[..., 0::2] = np.clip(box[..., 0::2] - x1, 0, img_w - 1)
        box_[..., 1::2] = np.clip(box[..., 1::2] - y1, 0, img_h - 1)
        return box_

    def _all_box_crop(self, results, crop_bbox):
        """Crop the gt_bboxes and proposals in results according to crop_bbox.

        Args:
            results (dict): All information about the sample, which contain
                'gt_bboxes' and 'proposals' (optional).
            crop_bbox(np.ndarray): The bbox used to crop the original image.
        """
        results['gt_bboxes'] = self._box_crop(results['gt_bboxes'], crop_bbox)
        if 'proposals' in results and results['proposals'] is not None:
            assert results['proposals'].shape[1] == 4
            results['proposals'] = self._box_crop(results['proposals'],
                                                  crop_bbox)
        return results

    def transform(self, results):
        """Performs the RandomCrop augmentation.

        Args:
            results (dict): The resulting dict to be modified and passed
                to the next transform in pipeline.
        """
        _init_lazy_if_proper(results, self.lazy)
        if 'keypoint' in results:
            assert not self.lazy, ('Keypoint Augmentations are not compatible '
                                   'with lazy == True')

        img_h, img_w = results['img_shape']
        assert self.size <= img_h and self.size <= img_w

        y_offset = 0
        x_offset = 0
        if img_h > self.size:
            y_offset = int(np.random.randint(0, img_h - self.size))
        if img_w > self.size:
            x_offset = int(np.random.randint(0, img_w - self.size))

        if 'crop_quadruple' not in results:
            results['crop_quadruple'] = np.array(
                [0, 0, 1, 1],  # x, y, w, h
                dtype=np.float32)

        x_ratio, y_ratio = x_offset / img_w, y_offset / img_h
        w_ratio, h_ratio = self.size / img_w, self.size / img_h

        old_crop_quadruple = results['crop_quadruple']
        old_x_ratio, old_y_ratio = old_crop_quadruple[0], old_crop_quadruple[1]
        old_w_ratio, old_h_ratio = old_crop_quadruple[2], old_crop_quadruple[3]
        new_crop_quadruple = [
            old_x_ratio + x_ratio * old_w_ratio,
            old_y_ratio + y_ratio * old_h_ratio, w_ratio * old_w_ratio,
            h_ratio * old_h_ratio
        ]
        results['crop_quadruple'] = np.array(
            new_crop_quadruple, dtype=np.float32)

        new_h, new_w = self.size, self.size

        crop_bbox = np.array(
            [x_offset, y_offset, x_offset + new_w, y_offset + new_h])
        results['crop_bbox'] = crop_bbox

        results['img_shape'] = (new_h, new_w)

        if not self.lazy:
            if 'keypoint' in results:
                results['keypoint'] = self._crop_kps(results['keypoint'],
                                                     crop_bbox)
            if 'imgs' in results:
                results['imgs'] = self._crop_imgs(results['imgs'], crop_bbox)
        else:
            lazyop = results['lazy']
            if lazyop['flip']:
                raise NotImplementedError('Put Flip at last for now')

            # record crop_bbox in lazyop dict to ensure only crop once in Fuse
            lazy_left, lazy_top, lazy_right, lazy_bottom = lazyop['crop_bbox']
            left = x_offset * (lazy_right - lazy_left) / img_w
            right = (x_offset + new_w) * (lazy_right - lazy_left) / img_w
            top = y_offset * (lazy_bottom - lazy_top) / img_h
            bottom = (y_offset + new_h) * (lazy_bottom - lazy_top) / img_h
            lazyop['crop_bbox'] = np.array([(lazy_left + left),
                                            (lazy_top + top),
                                            (lazy_left + right),
                                            (lazy_top + bottom)],
                                           dtype=np.float32)

        # Process entity boxes
        if 'gt_bboxes' in results:
            assert not self.lazy
            results = self._all_box_crop(results, results['crop_bbox'])

        return results

    def __repr__(self):
        repr_str = (f'{self.__class__.__name__}(size={self.size}, '
                    f'lazy={self.lazy})')
        return repr_str


@TRANSFORMS.register_module()
class RandomResizedCrop(RandomCrop):
    """Random crop that specifics the area and height-weight ratio range.

    Required keys in results are "img_shape", "crop_bbox", "imgs" (optional),
    "keypoint" (optional), added or modified keys are "imgs", "keypoint",
    "crop_bbox" and "lazy"; Required keys in "lazy" are "flip", "crop_bbox",
    added or modified key is "crop_bbox".

    Args:
        area_range (Tuple[float]): The candidate area scales range of
            output cropped images. Default: (0.08, 1.0).
        aspect_ratio_range (Tuple[float]): The candidate aspect ratio range of
            output cropped images. Default: (3 / 4, 4 / 3).
        lazy (bool): Determine whether to apply lazy operation. Default: False.
    """

    def __init__(self,
                 area_range=(0.08, 1.0),
                 aspect_ratio_range=(3 / 4, 4 / 3),
                 lazy=False):
        self.area_range = area_range
        self.aspect_ratio_range = aspect_ratio_range
        self.lazy = lazy
        if not mmengine.is_tuple_of(self.area_range, float):
            raise TypeError(f'Area_range must be a tuple of float, '
                            f'but got {type(area_range)}')
        if not mmengine.is_tuple_of(self.aspect_ratio_range, float):
            raise TypeError(f'Aspect_ratio_range must be a tuple of float, '
                            f'but got {type(aspect_ratio_range)}')

    @staticmethod
    def get_crop_bbox(img_shape,
                      area_range,
                      aspect_ratio_range,
                      max_attempts=10):
        """Get a crop bbox given the area range and aspect ratio range.

        Args:
            img_shape (Tuple[int]): Image shape
            area_range (Tuple[float]): The candidate area scales range of
                output cropped images. Default: (0.08, 1.0).
            aspect_ratio_range (Tuple[float]): The candidate aspect
                ratio range of output cropped images. Default: (3 / 4, 4 / 3).
                max_attempts (int): The maximum of attempts. Default: 10.
            max_attempts (int): Max attempts times to generate random candidate
                bounding box. If it doesn't qualified one, the center bounding
                box will be used.
        Returns:
            (list[int]) A random crop bbox within the area range and aspect
            ratio range.
        """
        assert 0 < area_range[0] <= area_range[1] <= 1
        assert 0 < aspect_ratio_range[0] <= aspect_ratio_range[1]

        img_h, img_w = img_shape
        area = img_h * img_w

        min_ar, max_ar = aspect_ratio_range
        aspect_ratios = np.exp(
            np.random.uniform(
                np.log(min_ar), np.log(max_ar), size=max_attempts))
        target_areas = np.random.uniform(*area_range, size=max_attempts) * area
        candidate_crop_w = np.round(np.sqrt(target_areas *
                                            aspect_ratios)).astype(np.int32)
        candidate_crop_h = np.round(np.sqrt(target_areas /
                                            aspect_ratios)).astype(np.int32)

        for i in range(max_attempts):
            crop_w = candidate_crop_w[i]
            crop_h = candidate_crop_h[i]
            if crop_h <= img_h and crop_w <= img_w:
                x_offset = random.randint(0, img_w - crop_w)
                y_offset = random.randint(0, img_h - crop_h)
                return x_offset, y_offset, x_offset + crop_w, y_offset + crop_h

        # Fallback
        crop_size = min(img_h, img_w)
        x_offset = (img_w - crop_size) // 2
        y_offset = (img_h - crop_size) // 2
        return x_offset, y_offset, x_offset + crop_size, y_offset + crop_size

    def transform(self, results):
        """Performs the RandomResizeCrop augmentation.

        Args:
            results (dict): The resulting dict to be modified and passed
                to the next transform in pipeline.
        """
        _init_lazy_if_proper(results, self.lazy)
        if 'keypoint' in results:
            assert not self.lazy, ('Keypoint Augmentations are not compatible '
                                   'with lazy == True')

        img_h, img_w = results['img_shape']

        left, top, right, bottom = self.get_crop_bbox(
            (img_h, img_w), self.area_range, self.aspect_ratio_range)
        new_h, new_w = bottom - top, right - left

        if 'crop_quadruple' not in results:
            results['crop_quadruple'] = np.array(
                [0, 0, 1, 1],  # x, y, w, h
                dtype=np.float32)

        x_ratio, y_ratio = left / img_w, top / img_h
        w_ratio, h_ratio = new_w / img_w, new_h / img_h

        old_crop_quadruple = results['crop_quadruple']
        old_x_ratio, old_y_ratio = old_crop_quadruple[0], old_crop_quadruple[1]
        old_w_ratio, old_h_ratio = old_crop_quadruple[2], old_crop_quadruple[3]
        new_crop_quadruple = [
            old_x_ratio + x_ratio * old_w_ratio,
            old_y_ratio + y_ratio * old_h_ratio, w_ratio * old_w_ratio,
            h_ratio * old_h_ratio
        ]
        results['crop_quadruple'] = np.array(
            new_crop_quadruple, dtype=np.float32)

        crop_bbox = np.array([left, top, right, bottom])
        results['crop_bbox'] = crop_bbox
        results['img_shape'] = (new_h, new_w)

        if not self.lazy:
            if 'keypoint' in results:
                results['keypoint'] = self._crop_kps(results['keypoint'],
                                                     crop_bbox)
            if 'imgs' in results:
                results['imgs'] = self._crop_imgs(results['imgs'], crop_bbox)
        else:
            lazyop = results['lazy']
            if lazyop['flip']:
                raise NotImplementedError('Put Flip at last for now')

            # record crop_bbox in lazyop dict to ensure only crop once in Fuse
            lazy_left, lazy_top, lazy_right, lazy_bottom = lazyop['crop_bbox']
            left = left * (lazy_right - lazy_left) / img_w
            right = right * (lazy_right - lazy_left) / img_w
            top = top * (lazy_bottom - lazy_top) / img_h
            bottom = bottom * (lazy_bottom - lazy_top) / img_h
            lazyop['crop_bbox'] = np.array([(lazy_left + left),
                                            (lazy_top + top),
                                            (lazy_left + right),
                                            (lazy_top + bottom)],
                                           dtype=np.float32)

        if 'gt_bboxes' in results:
            assert not self.lazy
            results = self._all_box_crop(results, results['crop_bbox'])

        return results

    def __repr__(self):
        repr_str = (f'{self.__class__.__name__}('
                    f'area_range={self.area_range}, '
                    f'aspect_ratio_range={self.aspect_ratio_range}, '
                    f'lazy={self.lazy})')
        return repr_str


@TRANSFORMS.register_module()
class MultiScaleCrop(RandomCrop):
    """Crop images with a list of randomly selected scales.

    Randomly select the w and h scales from a list of scales. Scale of 1 means
    the base size, which is the minimal of image width and height. The scale
    level of w and h is controlled to be smaller than a certain value to
    prevent too large or small aspect ratio.

    Required keys are "img_shape", "imgs" (optional), "keypoint" (optional),
    added or modified keys are "imgs", "crop_bbox", "img_shape", "lazy" and
    "scales". Required keys in "lazy" are "crop_bbox", added or modified key is
    "crop_bbox".

    Args:
        input_size (int | tuple[int]): (w, h) of network input.
        scales (tuple[float]): width and height scales to be selected.
        max_wh_scale_gap (int): Maximum gap of w and h scale levels.
            Default: 1.
        random_crop (bool): If set to True, the cropping bbox will be randomly
            sampled, otherwise it will be sampler from fixed regions.
            Default: False.
        num_fixed_crops (int): If set to 5, the cropping bbox will keep 5
            basic fixed regions: "upper left", "upper right", "lower left",
            "lower right", "center". If set to 13, the cropping bbox will
            append another 8 fix regions: "center left", "center right",
            "lower center", "upper center", "upper left quarter",
            "upper right quarter", "lower left quarter", "lower right quarter".
            Default: 5.
        lazy (bool): Determine whether to apply lazy operation. Default: False.
    """

    def __init__(self,
                 input_size,
                 scales=(1, ),
                 max_wh_scale_gap=1,
                 random_crop=False,
                 num_fixed_crops=5,
                 lazy=False):
        self.input_size = _pair(input_size)
        if not mmengine.is_tuple_of(self.input_size, int):
            raise TypeError(f'Input_size must be int or tuple of int, '
                            f'but got {type(input_size)}')

        if not isinstance(scales, tuple):
            raise TypeError(f'Scales must be tuple, but got {type(scales)}')

        if num_fixed_crops not in [5, 13]:
            raise ValueError(f'Num_fix_crops must be in {[5, 13]}, '
                             f'but got {num_fixed_crops}')

        self.scales = scales
        self.max_wh_scale_gap = max_wh_scale_gap
        self.random_crop = random_crop
        self.num_fixed_crops = num_fixed_crops
        self.lazy = lazy

    def transform(self, results):
        """Performs the MultiScaleCrop augmentation.

        Args:
            results (dict): The resulting dict to be modified and passed
                to the next transform in pipeline.
        """
        _init_lazy_if_proper(results, self.lazy)
        if 'keypoint' in results:
            assert not self.lazy, ('Keypoint Augmentations are not compatible '
                                   'with lazy == True')

        img_h, img_w = results['img_shape']
        base_size = min(img_h, img_w)
        crop_sizes = [int(base_size * s) for s in self.scales]

        candidate_sizes = []
        for i, h in enumerate(crop_sizes):
            for j, w in enumerate(crop_sizes):
                if abs(i - j) <= self.max_wh_scale_gap:
                    candidate_sizes.append([w, h])

        crop_size = random.choice(candidate_sizes)
        for i in range(2):
            if abs(crop_size[i] - self.input_size[i]) < 3:
                crop_size[i] = self.input_size[i]

        crop_w, crop_h = crop_size

        if self.random_crop:
            x_offset = random.randint(0, img_w - crop_w)
            y_offset = random.randint(0, img_h - crop_h)
        else:
            w_step = (img_w - crop_w) // 4
            h_step = (img_h - crop_h) // 4
            candidate_offsets = [
                (0, 0),  # upper left
                (4 * w_step, 0),  # upper right
                (0, 4 * h_step),  # lower left
                (4 * w_step, 4 * h_step),  # lower right
                (2 * w_step, 2 * h_step),  # center
            ]
            if self.num_fixed_crops == 13:
                extra_candidate_offsets = [
                    (0, 2 * h_step),  # center left
                    (4 * w_step, 2 * h_step),  # center right
                    (2 * w_step, 4 * h_step),  # lower center
                    (2 * w_step, 0 * h_step),  # upper center
                    (1 * w_step, 1 * h_step),  # upper left quarter
                    (3 * w_step, 1 * h_step),  # upper right quarter
                    (1 * w_step, 3 * h_step),  # lower left quarter
                    (3 * w_step, 3 * h_step)  # lower right quarter
                ]
                candidate_offsets.extend(extra_candidate_offsets)
            x_offset, y_offset = random.choice(candidate_offsets)

        new_h, new_w = crop_h, crop_w

        crop_bbox = np.array(
            [x_offset, y_offset, x_offset + new_w, y_offset + new_h])
        results['crop_bbox'] = crop_bbox
        results['img_shape'] = (new_h, new_w)
        results['scales'] = self.scales

        if 'crop_quadruple' not in results:
            results['crop_quadruple'] = np.array(
                [0, 0, 1, 1],  # x, y, w, h
                dtype=np.float32)

        x_ratio, y_ratio = x_offset / img_w, y_offset / img_h
        w_ratio, h_ratio = new_w / img_w, new_h / img_h

        old_crop_quadruple = results['crop_quadruple']
        old_x_ratio, old_y_ratio = old_crop_quadruple[0], old_crop_quadruple[1]
        old_w_ratio, old_h_ratio = old_crop_quadruple[2], old_crop_quadruple[3]
        new_crop_quadruple = [
            old_x_ratio + x_ratio * old_w_ratio,
            old_y_ratio + y_ratio * old_h_ratio, w_ratio * old_w_ratio,
            h_ratio * old_h_ratio
        ]
        results['crop_quadruple'] = np.array(
            new_crop_quadruple, dtype=np.float32)

        if not self.lazy:
            if 'keypoint' in results:
                results['keypoint'] = self._crop_kps(results['keypoint'],
                                                     crop_bbox)
            if 'imgs' in results:
                results['imgs'] = self._crop_imgs(results['imgs'], crop_bbox)
        else:
            lazyop = results['lazy']
            if lazyop['flip']:
                raise NotImplementedError('Put Flip at last for now')

            # record crop_bbox in lazyop dict to ensure only crop once in Fuse
            lazy_left, lazy_top, lazy_right, lazy_bottom = lazyop['crop_bbox']
            left = x_offset * (lazy_right - lazy_left) / img_w
            right = (x_offset + new_w) * (lazy_right - lazy_left) / img_w
            top = y_offset * (lazy_bottom - lazy_top) / img_h
            bottom = (y_offset + new_h) * (lazy_bottom - lazy_top) / img_h
            lazyop['crop_bbox'] = np.array([(lazy_left + left),
                                            (lazy_top + top),
                                            (lazy_left + right),
                                            (lazy_top + bottom)],
                                           dtype=np.float32)

        if 'gt_bboxes' in results:
            assert not self.lazy
            results = self._all_box_crop(results, results['crop_bbox'])

        return results

    def __repr__(self):
        repr_str = (f'{self.__class__.__name__}('
                    f'input_size={self.input_size}, scales={self.scales}, '
                    f'max_wh_scale_gap={self.max_wh_scale_gap}, '
                    f'random_crop={self.random_crop}, '
                    f'num_fixed_crops={self.num_fixed_crops}, '
                    f'lazy={self.lazy})')
        return repr_str


@TRANSFORMS.register_module()
class Resize(BaseTransform):
    """Resize images to a specific size.

    Required keys are "img_shape", "modality", "imgs" (optional), "keypoint"
    (optional), added or modified keys are "imgs", "img_shape", "keep_ratio",
    "scale_factor", "lazy", "resize_size". Required keys in "lazy" is None,
    added or modified key is "interpolation".

    Args:
        scale (float | Tuple[int]): If keep_ratio is True, it serves as scaling
            factor or maximum size:
            If it is a float number, the image will be rescaled by this
            factor, else if it is a tuple of 2 integers, the image will
            be rescaled as large as possible within the scale.
            Otherwise, it serves as (w, h) of output size.
        keep_ratio (bool): If set to True, Images will be resized without
            changing the aspect ratio. Otherwise, it will resize images to a
            given size. Default: True.
        interpolation (str): Algorithm used for interpolation,
            accepted values are "nearest", "bilinear", "bicubic", "area",
            "lanczos". Default: "bilinear".
        lazy (bool): Determine whether to apply lazy operation. Default: False.
    """

    def __init__(self,
                 scale,
                 keep_ratio=True,
                 interpolation='bilinear',
                 lazy=False):
        if isinstance(scale, float):
            if scale <= 0:
                raise ValueError(f'Invalid scale {scale}, must be positive.')
        elif isinstance(scale, tuple):
            max_long_edge = max(scale)
            max_short_edge = min(scale)
            if max_short_edge == -1:
                # assign np.inf to long edge for rescaling short edge later.
                scale = (np.inf, max_long_edge)
        else:
            raise TypeError(
                f'Scale must be float or tuple of int, but got {type(scale)}')
        self.scale = scale
        self.keep_ratio = keep_ratio
        self.interpolation = interpolation
        self.lazy = lazy

    def _resize_imgs(self, imgs, new_w, new_h):
        """Static method for resizing keypoint."""
        return [
            mmcv.imresize(
                img, (new_w, new_h), interpolation=self.interpolation)
            for img in imgs
        ]

    @staticmethod
    def _resize_kps(kps, scale_factor):
        """Static method for resizing keypoint."""
        return kps * scale_factor

    @staticmethod
    def _box_resize(box, scale_factor):
        """Rescale the bounding boxes according to the scale_factor.

        Args:
            box (np.ndarray): The bounding boxes.
            scale_factor (np.ndarray): The scale factor used for rescaling.
        """
        assert len(scale_factor) == 2
        scale_factor = np.concatenate([scale_factor, scale_factor])
        return box * scale_factor

    def transform(self, results):
        """Performs the Resize augmentation.

        Args:
            results (dict): The resulting dict to be modified and passed
                to the next transform in pipeline.
        """

        _init_lazy_if_proper(results, self.lazy)
        if 'keypoint' in results:
            assert not self.lazy, ('Keypoint Augmentations are not compatible '
                                   'with lazy == True')

        if 'scale_factor' not in results:
            results['scale_factor'] = np.array([1, 1], dtype=np.float32)
        img_h, img_w = results['img_shape']

        if self.keep_ratio:
            new_w, new_h = mmcv.rescale_size((img_w, img_h), self.scale)
        else:
            new_w, new_h = self.scale

        self.scale_factor = np.array([new_w / img_w, new_h / img_h],
                                     dtype=np.float32)

        results['img_shape'] = (new_h, new_w)
        results['keep_ratio'] = self.keep_ratio
        results['scale_factor'] = results['scale_factor'] * self.scale_factor

        if not self.lazy:
            if 'imgs' in results:
                results['imgs'] = self._resize_imgs(results['imgs'], new_w,
                                                    new_h)
            if 'keypoint' in results:
                results['keypoint'] = self._resize_kps(results['keypoint'],
                                                       self.scale_factor)
        else:
            lazyop = results['lazy']
            if lazyop['flip']:
                raise NotImplementedError('Put Flip at last for now')
            lazyop['interpolation'] = self.interpolation

        if 'gt_bboxes' in results:
            assert not self.lazy
            results['gt_bboxes'] = self._box_resize(results['gt_bboxes'],
                                                    self.scale_factor)
            if 'proposals' in results and results['proposals'] is not None:
                assert results['proposals'].shape[1] == 4
                results['proposals'] = self._box_resize(
                    results['proposals'], self.scale_factor)

        return results

    def __repr__(self):
        repr_str = (f'{self.__class__.__name__}('
                    f'scale={self.scale}, keep_ratio={self.keep_ratio}, '
                    f'interpolation={self.interpolation}, '
                    f'lazy={self.lazy})')
        return repr_str


@TRANSFORMS.register_module()
class RandomRescale(BaseTransform):
    """Randomly resize images so that the short_edge is resized to a specific
    size in a given range. The scale ratio is unchanged after resizing.

    Required keys are "imgs", "img_shape", "modality", added or modified
    keys are "imgs", "img_shape", "keep_ratio", "scale_factor", "resize_size",
    "short_edge".

    Args:
        scale_range (tuple[int]): The range of short edge length. A closed
            interval.
        interpolation (str): Algorithm used for interpolation:
            "nearest" | "bilinear". Default: "bilinear".
    """

    def __init__(self, scale_range, interpolation='bilinear'):
        self.scale_range = scale_range
        # make sure scale_range is legal, first make sure the type is OK
        assert mmengine.is_tuple_of(scale_range, int)
        assert len(scale_range) == 2
        assert scale_range[0] < scale_range[1]
        assert np.all([x > 0 for x in scale_range])

        self.keep_ratio = True
        self.interpolation = interpolation

    def transform(self, results):
        """Performs the Resize augmentation.

        Args:
            results (dict): The resulting dict to be modified and passed
                to the next transform in pipeline.
        """
        short_edge = np.random.randint(self.scale_range[0],
                                       self.scale_range[1] + 1)
        resize = Resize((-1, short_edge),
                        keep_ratio=True,
                        interpolation=self.interpolation,
                        lazy=False)
        results = resize(results)

        results['short_edge'] = short_edge
        return results

    def __repr__(self):
        scale_range = self.scale_range
        repr_str = (f'{self.__class__.__name__}('
                    f'scale_range=({scale_range[0]}, {scale_range[1]}), '
                    f'interpolation={self.interpolation})')
        return repr_str


@TRANSFORMS.register_module()
class Flip(BaseTransform):
    """Flip the input images with a probability.

    Reverse the order of elements in the given imgs with a specific direction.
    The shape of the imgs is preserved, but the elements are reordered.

    Required keys are "img_shape", "modality", "imgs" (optional), "keypoint"
    (optional), added or modified keys are "imgs", "keypoint", "lazy" and
    "flip_direction". Required keys in "lazy" is None, added or modified key
    are "flip" and "flip_direction". The Flip augmentation should be placed
    after any cropping / reshaping augmentations, to make sure crop_quadruple
    is calculated properly.

    Args:
        flip_ratio (float): Probability of implementing flip. Default: 0.5.
        direction (str): Flip imgs horizontally or vertically. Options are
            "horizontal" | "vertical". Default: "horizontal".
        flip_label_map (Dict[int, int] | None): Transform the label of the
            flipped image with the specific label. Default: None.
        left_kp (list[int]): Indexes of left keypoints, used to flip keypoints.
            Default: None.
        right_kp (list[ind]): Indexes of right keypoints, used to flip
            keypoints. Default: None.
        lazy (bool): Determine whether to apply lazy operation. Default: False.
    """
    _directions = ['horizontal', 'vertical']

    def __init__(self,
                 flip_ratio=0.5,
                 direction='horizontal',
                 flip_label_map=None,
                 left_kp=None,
                 right_kp=None,
                 lazy=False):
        if direction not in self._directions:
            raise ValueError(f'Direction {direction} is not supported. '
                             f'Currently support ones are {self._directions}')
        self.flip_ratio = flip_ratio
        self.direction = direction
        self.flip_label_map = flip_label_map
        self.left_kp = left_kp
        self.right_kp = right_kp
        self.lazy = lazy

    def _flip_imgs(self, imgs, modality):
        """Utility function for flipping images."""
        _ = [mmcv.imflip_(img, self.direction) for img in imgs]
        lt = len(imgs)
        if modality == 'Flow':
            # The 1st frame of each 2 frames is flow-x
            for i in range(0, lt, 2):
                imgs[i] = mmcv.iminvert(imgs[i])
        return imgs

    def _flip_kps(self, kps, kpscores, img_width):
        """Utility function for flipping keypoint."""
        kp_x = kps[..., 0]
        kp_x[kp_x != 0] = img_width - kp_x[kp_x != 0]
        new_order = list(range(kps.shape[2]))
        if self.left_kp is not None and self.right_kp is not None:
            for left, right in zip(self.left_kp, self.right_kp):
                new_order[left] = right
                new_order[right] = left
        kps = kps[:, :, new_order]
        if kpscores is not None:
            kpscores = kpscores[:, :, new_order]
        return kps, kpscores

    @staticmethod
    def _box_flip(box, img_width):
        """Flip the bounding boxes given the width of the image.

        Args:
            box (np.ndarray): The bounding boxes.
            img_width (int): The img width.
        """
        box_ = box.copy()
        box_[..., 0::4] = img_width - box[..., 2::4]
        box_[..., 2::4] = img_width - box[..., 0::4]
        return box_

    def transform(self, results):
        """Performs the Flip augmentation.

        Args:
            results (dict): The resulting dict to be modified and passed
                to the next transform in pipeline.
        """
        _init_lazy_if_proper(results, self.lazy)
        if 'keypoint' in results:
            assert not self.lazy, ('Keypoint Augmentations are not compatible '
                                   'with lazy == True')
            assert self.direction == 'horizontal', (
                'Only horizontal flips are'
                'supported for human keypoints')

        modality = results['modality']
        if modality == 'Flow':
            assert self.direction == 'horizontal'

        flip = np.random.rand() < self.flip_ratio

        results['flip'] = flip
        results['flip_direction'] = self.direction
        img_width = results['img_shape'][1]

        if self.flip_label_map is not None and flip:
            results['label'] = self.flip_label_map.get(results['label'],
                                                       results['label'])

        if not self.lazy:
            if flip:
                if 'imgs' in results:
                    results['imgs'] = self._flip_imgs(results['imgs'],
                                                      modality)
                if 'keypoint' in results:
                    kp = results['keypoint']
                    kpscore = results.get('keypoint_score', None)
                    kp, kpscore = self._flip_kps(kp, kpscore, img_width)
                    results['keypoint'] = kp
                    if 'keypoint_score' in results:
                        results['keypoint_score'] = kpscore
        else:
            lazyop = results['lazy']
            if lazyop['flip']:
                raise NotImplementedError('Use one Flip please')
            lazyop['flip'] = flip
            lazyop['flip_direction'] = self.direction

        if 'gt_bboxes' in results and flip:
            assert not self.lazy and self.direction == 'horizontal'
            width = results['img_shape'][1]
            results['gt_bboxes'] = self._box_flip(results['gt_bboxes'], width)
            if 'proposals' in results and results['proposals'] is not None:
                assert results['proposals'].shape[1] == 4
                results['proposals'] = self._box_flip(results['proposals'],
                                                      width)

        return results

    def __repr__(self):
        repr_str = (
            f'{self.__class__.__name__}('
            f'flip_ratio={self.flip_ratio}, direction={self.direction}, '
            f'flip_label_map={self.flip_label_map}, lazy={self.lazy})')
        return repr_str


@TRANSFORMS.register_module()
class ColorJitter(BaseTransform):
    """Perform ColorJitter to each img.

    Required keys are "imgs", added or modified keys are "imgs".

    Args:
        brightness (float | tuple[float]): The jitter range for brightness, if
            set as a float, the range will be (1 - brightness, 1 + brightness).
            Default: 0.5.
        contrast (float | tuple[float]): The jitter range for contrast, if set
            as a float, the range will be (1 - contrast, 1 + contrast).
            Default: 0.5.
        saturation (float | tuple[float]): The jitter range for saturation, if
            set as a float, the range will be (1 - saturation, 1 + saturation).
            Default: 0.5.
        hue (float | tuple[float]): The jitter range for hue, if set as a
            float, the range will be (-hue, hue). Default: 0.1.
    """

    @staticmethod
    def check_input(val, max, base):
        if isinstance(val, tuple):
            assert base - max <= val[0] <= val[1] <= base + max
            return val
        assert val <= max
        return (base - val, base + val)

    @staticmethod
    def rgb_to_grayscale(img):
        return 0.2989 * img[..., 0] + 0.587 * img[..., 1] + 0.114 * img[..., 2]

    @staticmethod
    def adjust_contrast(img, factor):
        val = np.mean(ColorJitter.rgb_to_grayscale(img))
        return factor * img + (1 - factor) * val

    @staticmethod
    def adjust_saturation(img, factor):
        gray = np.stack([ColorJitter.rgb_to_grayscale(img)] * 3, axis=-1)
        return factor * img + (1 - factor) * gray

    @staticmethod
    def adjust_hue(img, factor):
        img = np.clip(img, 0, 255).astype(np.uint8)
        hsv = cv2.cvtColor(img, cv2.COLOR_RGB2HSV)
        offset = int(factor * 255)
        hsv[..., 0] = (hsv[..., 0] + offset) % 180
        img = cv2.cvtColor(hsv, cv2.COLOR_HSV2RGB)
        return img.astype(np.float32)

    def __init__(self, brightness=0.5, contrast=0.5, saturation=0.5, hue=0.1):
        self.brightness = self.check_input(brightness, 1, 1)
        self.contrast = self.check_input(contrast, 1, 1)
        self.saturation = self.check_input(saturation, 1, 1)
        self.hue = self.check_input(hue, 0.5, 0)
        self.fn_idx = np.random.permutation(4)

    def transform(self, results):
        """Perform ColorJitter.

        Args:
            results (dict): The resulting dict to be modified and passed
                to the next transform in pipeline.
        """
        imgs = results['imgs']
        num_clips, clip_len = 1, len(imgs)

        new_imgs = []
        for i in range(num_clips):
            b = np.random.uniform(
                low=self.brightness[0], high=self.brightness[1])
            c = np.random.uniform(low=self.contrast[0], high=self.contrast[1])
            s = np.random.uniform(
                low=self.saturation[0], high=self.saturation[1])
            h = np.random.uniform(low=self.hue[0], high=self.hue[1])
            start, end = i * clip_len, (i + 1) * clip_len

            for img in imgs[start:end]:
                img = img.astype(np.float32)
                for fn_id in self.fn_idx:
                    if fn_id == 0 and b != 1:
                        img *= b
                    if fn_id == 1 and c != 1:
                        img = self.adjust_contrast(img, c)
                    if fn_id == 2 and s != 1:
                        img = self.adjust_saturation(img, s)
                    if fn_id == 3 and h != 0:
                        img = self.adjust_hue(img, h)
                img = np.clip(img, 0, 255).astype(np.uint8)
                new_imgs.append(img)
        results['imgs'] = new_imgs
        return results

    def __repr__(self):
        repr_str = (f'{self.__class__.__name__}('
                    f'brightness={self.brightness}, '
                    f'contrast={self.contrast}, '
                    f'saturation={self.saturation}, '
                    f'hue={self.hue})')
        return repr_str


@TRANSFORMS.register_module()
class CenterCrop(RandomCrop):
    """Crop the center area from images.

    Required keys are "img_shape", "imgs" (optional), "keypoint" (optional),
    added or modified keys are "imgs", "keypoint", "crop_bbox", "lazy" and
    "img_shape". Required keys in "lazy" is "crop_bbox", added or modified key
    is "crop_bbox".

    Args:
        crop_size (int | tuple[int]): (w, h) of crop size.
        lazy (bool): Determine whether to apply lazy operation. Default: False.
    """

    def __init__(self, crop_size, lazy=False):
        self.crop_size = _pair(crop_size)
        self.lazy = lazy
        if not mmengine.is_tuple_of(self.crop_size, int):
            raise TypeError(f'Crop_size must be int or tuple of int, '
                            f'but got {type(crop_size)}')

    def transform(self, results):
        """Performs the CenterCrop augmentation.

        Args:
            results (dict): The resulting dict to be modified and passed
                to the next transform in pipeline.
        """
        _init_lazy_if_proper(results, self.lazy)
        if 'keypoint' in results:
            assert not self.lazy, ('Keypoint Augmentations are not compatible '
                                   'with lazy == True')

        img_h, img_w = results['img_shape']
        crop_w, crop_h = self.crop_size

        left = (img_w - crop_w) // 2
        top = (img_h - crop_h) // 2
        right = left + crop_w
        bottom = top + crop_h
        new_h, new_w = bottom - top, right - left

        crop_bbox = np.array([left, top, right, bottom])
        results['crop_bbox'] = crop_bbox
        results['img_shape'] = (new_h, new_w)

        if 'crop_quadruple' not in results:
            results['crop_quadruple'] = np.array(
                [0, 0, 1, 1],  # x, y, w, h
                dtype=np.float32)

        x_ratio, y_ratio = left / img_w, top / img_h
        w_ratio, h_ratio = new_w / img_w, new_h / img_h

        old_crop_quadruple = results['crop_quadruple']
        old_x_ratio, old_y_ratio = old_crop_quadruple[0], old_crop_quadruple[1]
        old_w_ratio, old_h_ratio = old_crop_quadruple[2], old_crop_quadruple[3]
        new_crop_quadruple = [
            old_x_ratio + x_ratio * old_w_ratio,
            old_y_ratio + y_ratio * old_h_ratio, w_ratio * old_w_ratio,
            h_ratio * old_h_ratio
        ]
        results['crop_quadruple'] = np.array(
            new_crop_quadruple, dtype=np.float32)

        if not self.lazy:
            if 'keypoint' in results:
                results['keypoint'] = self._crop_kps(results['keypoint'],
                                                     crop_bbox)
            if 'imgs' in results:
                results['imgs'] = self._crop_imgs(results['imgs'], crop_bbox)
        else:
            lazyop = results['lazy']
            if lazyop['flip']:
                raise NotImplementedError('Put Flip at last for now')

            # record crop_bbox in lazyop dict to ensure only crop once in Fuse
            lazy_left, lazy_top, lazy_right, lazy_bottom = lazyop['crop_bbox']
            left = left * (lazy_right - lazy_left) / img_w
            right = right * (lazy_right - lazy_left) / img_w
            top = top * (lazy_bottom - lazy_top) / img_h
            bottom = bottom * (lazy_bottom - lazy_top) / img_h
            lazyop['crop_bbox'] = np.array([(lazy_left + left),
                                            (lazy_top + top),
                                            (lazy_left + right),
                                            (lazy_top + bottom)],
                                           dtype=np.float32)

        if 'gt_bboxes' in results:
            assert not self.lazy
            results = self._all_box_crop(results, results['crop_bbox'])

        return results

    def __repr__(self):
        repr_str = (f'{self.__class__.__name__}(crop_size={self.crop_size}, '
                    f'lazy={self.lazy})')
        return repr_str


@TRANSFORMS.register_module()
class ThreeCrop(BaseTransform):
    """Crop images into three crops.

    Crop the images equally into three crops with equal intervals along the
    shorter side.
    Required keys are "imgs", "img_shape", added or modified keys are "imgs",
    "crop_bbox" and "img_shape".

    Args:
        crop_size(int | tuple[int]): (w, h) of crop size.
    """

    def __init__(self, crop_size):
        self.crop_size = _pair(crop_size)
        if not mmengine.is_tuple_of(self.crop_size, int):
            raise TypeError(f'Crop_size must be int or tuple of int, '
                            f'but got {type(crop_size)}')

    def transform(self, results):
        """Performs the ThreeCrop augmentation.

        Args:
            results (dict): The resulting dict to be modified and passed
                to the next transform in pipeline.
        """
        _init_lazy_if_proper(results, False)
        if 'gt_bboxes' in results or 'proposals' in results:
            warnings.warn('ThreeCrop cannot process bounding boxes')

        imgs = results['imgs']
        img_h, img_w = results['imgs'][0].shape[:2]
        crop_w, crop_h = self.crop_size
        assert crop_h == img_h or crop_w == img_w

        if crop_h == img_h:
            w_step = (img_w - crop_w) // 2
            offsets = [
                (0, 0),  # left
                (2 * w_step, 0),  # right
                (w_step, 0),  # middle
            ]
        elif crop_w == img_w:
            h_step = (img_h - crop_h) // 2
            offsets = [
                (0, 0),  # top
                (0, 2 * h_step),  # down
                (0, h_step),  # middle
            ]

        cropped = []
        crop_bboxes = []
        for x_offset, y_offset in offsets:
            bbox = [x_offset, y_offset, x_offset + crop_w, y_offset + crop_h]
            crop = [
                img[y_offset:y_offset + crop_h, x_offset:x_offset + crop_w]
                for img in imgs
            ]
            cropped.extend(crop)
            crop_bboxes.extend([bbox for _ in range(len(imgs))])

        crop_bboxes = np.array(crop_bboxes)
        results['imgs'] = cropped
        results['crop_bbox'] = crop_bboxes
        results['img_shape'] = results['imgs'][0].shape[:2]

        return results

    def __repr__(self):
        repr_str = f'{self.__class__.__name__}(crop_size={self.crop_size})'
        return repr_str


@TRANSFORMS.register_module()
class TenCrop(BaseTransform):
    """Crop the images into 10 crops (corner + center + flip).

    Crop the four corners and the center part of the image with the same
    given crop_size, and flip it horizontally.
    Required keys are "imgs", "img_shape", added or modified keys are "imgs",
    "crop_bbox" and "img_shape".

    Args:
        crop_size(int | tuple[int]): (w, h) of crop size.
    """

    def __init__(self, crop_size):
        self.crop_size = _pair(crop_size)
        if not mmengine.is_tuple_of(self.crop_size, int):
            raise TypeError(f'Crop_size must be int or tuple of int, '
                            f'but got {type(crop_size)}')

    def transform(self, results):
        """Performs the TenCrop augmentation.

        Args:
            results (dict): The resulting dict to be modified and passed
                to the next transform in pipeline.
        """
        _init_lazy_if_proper(results, False)

        if 'gt_bboxes' in results or 'proposals' in results:
            warnings.warn('TenCrop cannot process bounding boxes')

        imgs = results['imgs']

        img_h, img_w = results['imgs'][0].shape[:2]
        crop_w, crop_h = self.crop_size

        w_step = (img_w - crop_w) // 4
        h_step = (img_h - crop_h) // 4

        offsets = [
            (0, 0),  # upper left
            (4 * w_step, 0),  # upper right
            (0, 4 * h_step),  # lower left
            (4 * w_step, 4 * h_step),  # lower right
            (2 * w_step, 2 * h_step),  # center
        ]

        img_crops = list()
        crop_bboxes = list()
        for x_offset, y_offsets in offsets:
            crop = [
                img[y_offsets:y_offsets + crop_h, x_offset:x_offset + crop_w]
                for img in imgs
            ]
            flip_crop = [np.flip(c, axis=1).copy() for c in crop]
            bbox = [x_offset, y_offsets, x_offset + crop_w, y_offsets + crop_h]
            img_crops.extend(crop)
            img_crops.extend(flip_crop)
            crop_bboxes.extend([bbox for _ in range(len(imgs) * 2)])

        crop_bboxes = np.array(crop_bboxes)
        results['imgs'] = img_crops
        results['crop_bbox'] = crop_bboxes
        results['img_shape'] = results['imgs'][0].shape[:2]

        return results

    def __repr__(self):
        repr_str = f'{self.__class__.__name__}(crop_size={self.crop_size})'
        return repr_str


@TRANSFORMS.register_module()
class RandomErasing(BaseTransform):
    """Randomly selects a rectangle region in an image and erase pixels.
    basically refer mmcls.

    **Required Keys:**

    - img

    **Modified Keys:**

    - img

    Args:
        erase_prob (float): Probability that image will be randomly erased.
            Default: 0.5
        min_area_ratio (float): Minimum erased area / input image area
            Default: 0.02
        max_area_ratio (float): Maximum erased area / input image area
            Default: 1/3
        aspect_range (sequence | float): Aspect ratio range of erased area.
            if float, it will be converted to (aspect_ratio, 1/aspect_ratio)
            Default: (3/10, 10/3)
        mode (str): Fill method in erased area, can be:

            - const (default): All pixels are assign with the same value.
            - rand: each pixel is assigned with a random value in [0, 255]

        fill_color (sequence | Number): Base color filled in erased area.
            Defaults to (128, 128, 128).
        fill_std (sequence | Number, optional): If set and ``mode`` is 'rand',
            fill erased area with random color from normal distribution
            (mean=fill_color, std=fill_std); If not set, fill erased area with
            random color from uniform distribution (0~255). Defaults to None.

    Note:
        See `Random Erasing Data Augmentation
        <https://arxiv.org/pdf/1708.04896.pdf>`_

        This paper provided 4 modes: RE-R, RE-M, RE-0, RE-255, and use RE-M as
        default. The config of these 4 modes are:

        - RE-R: RandomErasing(mode='rand')
        - RE-M: RandomErasing(mode='const', fill_color=(123.67, 116.3, 103.5))
        - RE-0: RandomErasing(mode='const', fill_color=0)
        - RE-255: RandomErasing(mode='const', fill_color=255)
    """

    def __init__(self,
                 erase_prob=0.5,
                 min_area_ratio=0.02,
                 max_area_ratio=1 / 3,
                 aspect_range=(3 / 10, 10 / 3),
                 mode='const',
                 fill_color=(128, 128, 128),
                 fill_std=None):
        assert isinstance(erase_prob, float) and 0. <= erase_prob <= 1.
        assert isinstance(min_area_ratio, float) and 0. <= min_area_ratio <= 1.
        assert isinstance(max_area_ratio, float) and 0. <= max_area_ratio <= 1.
        assert min_area_ratio <= max_area_ratio, \
            'min_area_ratio should be smaller than max_area_ratio'
        if isinstance(aspect_range, float):
            aspect_range = min(aspect_range, 1 / aspect_range)
            aspect_range = (aspect_range, 1 / aspect_range)
        assert isinstance(aspect_range, Sequence) and len(aspect_range) == 2 \
            and all(isinstance(x, float) for x in aspect_range), \
            'aspect_range should be a float or Sequence with two float.'
        assert all(x > 0 for x in aspect_range), \
            'aspect_range should be positive.'
        assert aspect_range[0] <= aspect_range[1], \
            'In aspect_range (min, max), min should be smaller than max.'
        assert mode in ['const', 'rand'], \
            'Please select `mode` from ["const", "rand"].'
        if isinstance(fill_color, Number):
            fill_color = [fill_color] * 3
        assert isinstance(fill_color, Sequence) and len(fill_color) == 3 \
            and all(isinstance(x, Number) for x in fill_color), \
            'fill_color should be a float or Sequence with three int.'
        if fill_std is not None:
            if isinstance(fill_std, Number):
                fill_std = [fill_std] * 3
            assert isinstance(fill_std, Sequence) and len(fill_std) == 3 \
                and all(isinstance(x, Number) for x in fill_std), \
                'fill_std should be a float or Sequence with three int.'

        self.erase_prob = erase_prob
        self.min_area_ratio = min_area_ratio
        self.max_area_ratio = max_area_ratio
        self.aspect_range = aspect_range
        self.mode = mode
        self.fill_color = fill_color
        self.fill_std = fill_std

    def _img_fill_pixels(self, img, top, left, h, w):
        """Fill pixels to the patch of image."""
        if self.mode == 'const':
            patch = np.empty((h, w, 3), dtype=np.uint8)
            patch[:, :] = np.array(self.fill_color, dtype=np.uint8)
        elif self.fill_std is None:
            # Uniform distribution
            patch = np.random.uniform(0, 256, (h, w, 3)).astype(np.uint8)
        else:
            # Normal distribution
            patch = np.random.normal(self.fill_color, self.fill_std, (h, w, 3))
            patch = np.clip(patch.astype(np.int32), 0, 255).astype(np.uint8)

        img[top:top + h, left:left + w] = patch
        return img

    def _fill_pixels(self, imgs, top, left, h, w):
        """Fill pixels to the patch of each image in frame clip."""
        return [self._img_fill_pixels(img, top, left, h, w) for img in imgs]

    @cache_randomness
    def random_disable(self):
        """Randomly disable the transform."""
        return np.random.rand() > self.erase_prob

    @cache_randomness
    def random_patch(self, img_h, img_w):
        """Randomly generate patch the erase."""
        # convert the aspect ratio to log space to equally handle width and
        # height.
        log_aspect_range = np.log(
            np.array(self.aspect_range, dtype=np.float32))
        aspect_ratio = np.exp(np.random.uniform(*log_aspect_range))
        area = img_h * img_w
        area *= np.random.uniform(self.min_area_ratio, self.max_area_ratio)

        h = min(int(round(np.sqrt(area * aspect_ratio))), img_h)
        w = min(int(round(np.sqrt(area / aspect_ratio))), img_w)
        top = np.random.randint(0, img_h - h) if img_h > h else 0
        left = np.random.randint(0, img_w - w) if img_w > w else 0
        return top, left, h, w

    def transform(self, results):
        """
        Args:
            results (dict): Results dict from pipeline

        Returns:
            dict: Results after the transformation.
        """
        if self.random_disable():
            return results

        imgs = results['imgs']
        img_h, img_w = imgs[0].shape[:2]

        imgs = self._fill_pixels(imgs, *self.random_patch(img_h, img_w))

        results['imgs'] = imgs

        return results

    def __repr__(self):
        repr_str = self.__class__.__name__
        repr_str += f'(erase_prob={self.erase_prob}, '
        repr_str += f'min_area_ratio={self.min_area_ratio}, '
        repr_str += f'max_area_ratio={self.max_area_ratio}, '
        repr_str += f'aspect_range={self.aspect_range}, '
        repr_str += f'mode={self.mode}, '
        repr_str += f'fill_color={self.fill_color}, '
        repr_str += f'fill_std={self.fill_std})'
        return repr_str