Reference for ultralytics/data/augment.py

class BaseTransform:

    def __init__(self) -> None:
        pass

    def apply_image(self, labels):
        """Applies image transformation to labels."""
        pass

    def apply_instances(self, labels):
        """Applies transformations to input 'labels' and returns object instances."""
        pass

    def apply_semantic(self, labels):
        """Applies semantic segmentation to an image."""
        pass

    def __call__(self, labels):
        """Applies label transformations to an image, instances and semantic masks."""
        self.apply_image(labels)
        self.apply_instances(labels)
        self.apply_semantic(labels)

class Compose:

    def __init__(self, transforms):
        """Initializes the Compose object with a list of transforms."""
        self.transforms = transforms

    def __call__(self, data):
        """Applies a series of transformations to input data."""
        for t in self.transforms:
            data = t(data)
        return data

    def append(self, transform):
        """Appends a new transform to the existing list of transforms."""
        self.transforms.append(transform)

    def tolist(self):
        """Converts list of transforms to a standard Python list."""
        return self.transforms

    def __repr__(self):
        """Return string representation of object."""
        return f"{self.__class__.__name__}({', '.join([f'{t}' for t in self.transforms])})"

class BaseMixTransform:
    """This implementation is from mmyolo."""

    def __init__(self, dataset, pre_transform=None, p=0.0) -> None:
        self.dataset = dataset
        self.pre_transform = pre_transform
        self.p = p

    def __call__(self, labels):
        """Applies pre-processing transforms and mixup/mosaic transforms to labels data."""
        if random.uniform(0, 1) > self.p:
            return labels

        # Get index of one or three other images
        indexes = self.get_indexes()
        if isinstance(indexes, int):
            indexes = [indexes]

        # Get images information will be used for Mosaic or MixUp
        mix_labels = [self.dataset.get_image_and_label(i) for i in indexes]

        if self.pre_transform is not None:
            for i, data in enumerate(mix_labels):
                mix_labels[i] = self.pre_transform(data)
        labels['mix_labels'] = mix_labels

        # Mosaic or MixUp
        labels = self._mix_transform(labels)
        labels.pop('mix_labels', None)
        return labels

    def _mix_transform(self, labels):
        """Applies MixUp or Mosaic augmentation to the label dictionary."""
        raise NotImplementedError

    def get_indexes(self):
        """Gets a list of shuffled indexes for mosaic augmentation."""
        raise NotImplementedError

class Mosaic(BaseMixTransform):
    """
    Mosaic augmentation.

    This class performs mosaic augmentation by combining multiple (4 or 9) images into a single mosaic image.
    The augmentation is applied to a dataset with a given probability.

    Attributes:
        dataset: The dataset on which the mosaic augmentation is applied.
        imgsz (int, optional): Image size (height and width) after mosaic pipeline of a single image. Default to 640.
        p (float, optional): Probability of applying the mosaic augmentation. Must be in the range 0-1. Default to 1.0.
        n (int, optional): The grid size, either 4 (for 2x2) or 9 (for 3x3).
    """

    def __init__(self, dataset, imgsz=640, p=1.0, n=4):
        """Initializes the object with a dataset, image size, probability, and border."""
        assert 0 <= p <= 1.0, f'The probability should be in range [0, 1], but got {p}.'
        assert n in (4, 9), 'grid must be equal to 4 or 9.'
        super().__init__(dataset=dataset, p=p)
        self.dataset = dataset
        self.imgsz = imgsz
        self.border = (-imgsz // 2, -imgsz // 2)  # width, height
        self.n = n

    def get_indexes(self, buffer=True):
        """Return a list of random indexes from the dataset."""
        if buffer:  # select images from buffer
            return random.choices(list(self.dataset.buffer), k=self.n - 1)
        else:  # select any images
            return [random.randint(0, len(self.dataset) - 1) for _ in range(self.n - 1)]

    def _mix_transform(self, labels):
        """Apply mixup transformation to the input image and labels."""
        assert labels.get('rect_shape', None) is None, 'rect and mosaic are mutually exclusive.'
        assert len(labels.get('mix_labels', [])), 'There are no other images for mosaic augment.'
        return self._mosaic4(labels) if self.n == 4 else self._mosaic9(labels)

    def _mosaic4(self, labels):
        """Create a 2x2 image mosaic."""
        mosaic_labels = []
        s = self.imgsz
        yc, xc = (int(random.uniform(-x, 2 * s + x)) for x in self.border)  # mosaic center x, y
        for i in range(4):
            labels_patch = labels if i == 0 else labels['mix_labels'][i - 1]
            # Load image
            img = labels_patch['img']
            h, w = labels_patch.pop('resized_shape')

            # Place img in img4
            if i == 0:  # top left
                img4 = np.full((s * 2, s * 2, img.shape[2]), 114, dtype=np.uint8)  # base image with 4 tiles
                x1a, y1a, x2a, y2a = max(xc - w, 0), max(yc - h, 0), xc, yc  # xmin, ymin, xmax, ymax (large image)
                x1b, y1b, x2b, y2b = w - (x2a - x1a), h - (y2a - y1a), w, h  # xmin, ymin, xmax, ymax (small image)
            elif i == 1:  # top right
                x1a, y1a, x2a, y2a = xc, max(yc - h, 0), min(xc + w, s * 2), yc
                x1b, y1b, x2b, y2b = 0, h - (y2a - y1a), min(w, x2a - x1a), h
            elif i == 2:  # bottom left
                x1a, y1a, x2a, y2a = max(xc - w, 0), yc, xc, min(s * 2, yc + h)
                x1b, y1b, x2b, y2b = w - (x2a - x1a), 0, w, min(y2a - y1a, h)
            elif i == 3:  # bottom right
                x1a, y1a, x2a, y2a = xc, yc, min(xc + w, s * 2), min(s * 2, yc + h)
                x1b, y1b, x2b, y2b = 0, 0, min(w, x2a - x1a), min(y2a - y1a, h)

            img4[y1a:y2a, x1a:x2a] = img[y1b:y2b, x1b:x2b]  # img4[ymin:ymax, xmin:xmax]
            padw = x1a - x1b
            padh = y1a - y1b

            labels_patch = self._update_labels(labels_patch, padw, padh)
            mosaic_labels.append(labels_patch)
        final_labels = self._cat_labels(mosaic_labels)
        final_labels['img'] = img4
        return final_labels

    def _mosaic9(self, labels):
        """Create a 3x3 image mosaic."""
        mosaic_labels = []
        s = self.imgsz
        hp, wp = -1, -1  # height, width previous
        for i in range(9):
            labels_patch = labels if i == 0 else labels['mix_labels'][i - 1]
            # Load image
            img = labels_patch['img']
            h, w = labels_patch.pop('resized_shape')

            # Place img in img9
            if i == 0:  # center
                img9 = np.full((s * 3, s * 3, img.shape[2]), 114, dtype=np.uint8)  # base image with 4 tiles
                h0, w0 = h, w
                c = s, s, s + w, s + h  # xmin, ymin, xmax, ymax (base) coordinates
            elif i == 1:  # top
                c = s, s - h, s + w, s
            elif i == 2:  # top right
                c = s + wp, s - h, s + wp + w, s
            elif i == 3:  # right
                c = s + w0, s, s + w0 + w, s + h
            elif i == 4:  # bottom right
                c = s + w0, s + hp, s + w0 + w, s + hp + h
            elif i == 5:  # bottom
                c = s + w0 - w, s + h0, s + w0, s + h0 + h
            elif i == 6:  # bottom left
                c = s + w0 - wp - w, s + h0, s + w0 - wp, s + h0 + h
            elif i == 7:  # left
                c = s - w, s + h0 - h, s, s + h0
            elif i == 8:  # top left
                c = s - w, s + h0 - hp - h, s, s + h0 - hp

            padw, padh = c[:2]
            x1, y1, x2, y2 = (max(x, 0) for x in c)  # allocate coords

            # Image
            img9[y1:y2, x1:x2] = img[y1 - padh:, x1 - padw:]  # img9[ymin:ymax, xmin:xmax]
            hp, wp = h, w  # height, width previous for next iteration

            # Labels assuming imgsz*2 mosaic size
            labels_patch = self._update_labels(labels_patch, padw + self.border[0], padh + self.border[1])
            mosaic_labels.append(labels_patch)
        final_labels = self._cat_labels(mosaic_labels)

        final_labels['img'] = img9[-self.border[0]:self.border[0], -self.border[1]:self.border[1]]
        return final_labels

    @staticmethod
    def _update_labels(labels, padw, padh):
        """Update labels."""
        nh, nw = labels['img'].shape[:2]
        labels['instances'].convert_bbox(format='xyxy')
        labels['instances'].denormalize(nw, nh)
        labels['instances'].add_padding(padw, padh)
        return labels

    def _cat_labels(self, mosaic_labels):
        """Return labels with mosaic border instances clipped."""
        if len(mosaic_labels) == 0:
            return {}
        cls = []
        instances = []
        imgsz = self.imgsz * 2  # mosaic imgsz
        for labels in mosaic_labels:
            cls.append(labels['cls'])
            instances.append(labels['instances'])
        final_labels = {
            'im_file': mosaic_labels[0]['im_file'],
            'ori_shape': mosaic_labels[0]['ori_shape'],
            'resized_shape': (imgsz, imgsz),
            'cls': np.concatenate(cls, 0),
            'instances': Instances.concatenate(instances, axis=0),
            'mosaic_border': self.border}  # final_labels
        final_labels['instances'].clip(imgsz, imgsz)
        good = final_labels['instances'].remove_zero_area_boxes()
        final_labels['cls'] = final_labels['cls'][good]
        return final_labels

class MixUp(BaseMixTransform):

    def __init__(self, dataset, pre_transform=None, p=0.0) -> None:
        super().__init__(dataset=dataset, pre_transform=pre_transform, p=p)

    def get_indexes(self):
        """Get a random index from the dataset."""
        return random.randint(0, len(self.dataset) - 1)

    def _mix_transform(self, labels):
        """Applies MixUp augmentation https://arxiv.org/pdf/1710.09412.pdf."""
        r = np.random.beta(32.0, 32.0)  # mixup ratio, alpha=beta=32.0
        labels2 = labels['mix_labels'][0]
        labels['img'] = (labels['img'] * r + labels2['img'] * (1 - r)).astype(np.uint8)
        labels['instances'] = Instances.concatenate([labels['instances'], labels2['instances']], axis=0)
        labels['cls'] = np.concatenate([labels['cls'], labels2['cls']], 0)
        return labels

class RandomPerspective:

    def __init__(self,
                 degrees=0.0,
                 translate=0.1,
                 scale=0.5,
                 shear=0.0,
                 perspective=0.0,
                 border=(0, 0),
                 pre_transform=None):
        self.degrees = degrees
        self.translate = translate
        self.scale = scale
        self.shear = shear
        self.perspective = perspective
        # Mosaic border
        self.border = border
        self.pre_transform = pre_transform

    def affine_transform(self, img, border):
        """Center."""
        C = np.eye(3, dtype=np.float32)

        C[0, 2] = -img.shape[1] / 2  # x translation (pixels)
        C[1, 2] = -img.shape[0] / 2  # y translation (pixels)

        # Perspective
        P = np.eye(3, dtype=np.float32)
        P[2, 0] = random.uniform(-self.perspective, self.perspective)  # x perspective (about y)
        P[2, 1] = random.uniform(-self.perspective, self.perspective)  # y perspective (about x)

        # Rotation and Scale
        R = np.eye(3, dtype=np.float32)
        a = random.uniform(-self.degrees, self.degrees)
        # a += random.choice([-180, -90, 0, 90])  # add 90deg rotations to small rotations
        s = random.uniform(1 - self.scale, 1 + self.scale)
        # s = 2 ** random.uniform(-scale, scale)
        R[:2] = cv2.getRotationMatrix2D(angle=a, center=(0, 0), scale=s)

        # Shear
        S = np.eye(3, dtype=np.float32)
        S[0, 1] = math.tan(random.uniform(-self.shear, self.shear) * math.pi / 180)  # x shear (deg)
        S[1, 0] = math.tan(random.uniform(-self.shear, self.shear) * math.pi / 180)  # y shear (deg)

        # Translation
        T = np.eye(3, dtype=np.float32)
        T[0, 2] = random.uniform(0.5 - self.translate, 0.5 + self.translate) * self.size[0]  # x translation (pixels)
        T[1, 2] = random.uniform(0.5 - self.translate, 0.5 + self.translate) * self.size[1]  # y translation (pixels)

        # Combined rotation matrix
        M = T @ S @ R @ P @ C  # order of operations (right to left) is IMPORTANT
        # Affine image
        if (border[0] != 0) or (border[1] != 0) or (M != np.eye(3)).any():  # image changed
            if self.perspective:
                img = cv2.warpPerspective(img, M, dsize=self.size, borderValue=(114, 114, 114))
            else:  # affine
                img = cv2.warpAffine(img, M[:2], dsize=self.size, borderValue=(114, 114, 114))
        return img, M, s

    def apply_bboxes(self, bboxes, M):
        """
        Apply affine to bboxes only.

        Args:
            bboxes (ndarray): list of bboxes, xyxy format, with shape (num_bboxes, 4).
            M (ndarray): affine matrix.

        Returns:
            new_bboxes (ndarray): bboxes after affine, [num_bboxes, 4].
        """
        n = len(bboxes)
        if n == 0:
            return bboxes

        xy = np.ones((n * 4, 3), dtype=bboxes.dtype)
        xy[:, :2] = bboxes[:, [0, 1, 2, 3, 0, 3, 2, 1]].reshape(n * 4, 2)  # x1y1, x2y2, x1y2, x2y1
        xy = xy @ M.T  # transform
        xy = (xy[:, :2] / xy[:, 2:3] if self.perspective else xy[:, :2]).reshape(n, 8)  # perspective rescale or affine

        # Create new boxes
        x = xy[:, [0, 2, 4, 6]]
        y = xy[:, [1, 3, 5, 7]]
        return np.concatenate((x.min(1), y.min(1), x.max(1), y.max(1)), dtype=bboxes.dtype).reshape(4, n).T

    def apply_segments(self, segments, M):
        """
        Apply affine to segments and generate new bboxes from segments.

        Args:
            segments (ndarray): list of segments, [num_samples, 500, 2].
            M (ndarray): affine matrix.

        Returns:
            new_segments (ndarray): list of segments after affine, [num_samples, 500, 2].
            new_bboxes (ndarray): bboxes after affine, [N, 4].
        """
        n, num = segments.shape[:2]
        if n == 0:
            return [], segments

        xy = np.ones((n * num, 3), dtype=segments.dtype)
        segments = segments.reshape(-1, 2)
        xy[:, :2] = segments
        xy = xy @ M.T  # transform
        xy = xy[:, :2] / xy[:, 2:3]
        segments = xy.reshape(n, -1, 2)
        bboxes = np.stack([segment2box(xy, self.size[0], self.size[1]) for xy in segments], 0)
        return bboxes, segments

    def apply_keypoints(self, keypoints, M):
        """
        Apply affine to keypoints.

        Args:
            keypoints (ndarray): keypoints, [N, 17, 3].
            M (ndarray): affine matrix.

        Returns:
            new_keypoints (ndarray): keypoints after affine, [N, 17, 3].
        """
        n, nkpt = keypoints.shape[:2]
        if n == 0:
            return keypoints
        xy = np.ones((n * nkpt, 3), dtype=keypoints.dtype)
        visible = keypoints[..., 2].reshape(n * nkpt, 1)
        xy[:, :2] = keypoints[..., :2].reshape(n * nkpt, 2)
        xy = xy @ M.T  # transform
        xy = xy[:, :2] / xy[:, 2:3]  # perspective rescale or affine
        out_mask = (xy[:, 0] < 0) | (xy[:, 1] < 0) | (xy[:, 0] > self.size[0]) | (xy[:, 1] > self.size[1])
        visible[out_mask] = 0
        return np.concatenate([xy, visible], axis=-1).reshape(n, nkpt, 3)

    def __call__(self, labels):
        """
        Affine images and targets.

        Args:
            labels (dict): a dict of `bboxes`, `segments`, `keypoints`.
        """
        if self.pre_transform and 'mosaic_border' not in labels:
            labels = self.pre_transform(labels)
        labels.pop('ratio_pad', None)  # do not need ratio pad

        img = labels['img']
        cls = labels['cls']
        instances = labels.pop('instances')
        # Make sure the coord formats are right
        instances.convert_bbox(format='xyxy')
        instances.denormalize(*img.shape[:2][::-1])

        border = labels.pop('mosaic_border', self.border)
        self.size = img.shape[1] + border[1] * 2, img.shape[0] + border[0] * 2  # w, h
        # M is affine matrix
        # scale for func:`box_candidates`
        img, M, scale = self.affine_transform(img, border)

        bboxes = self.apply_bboxes(instances.bboxes, M)

        segments = instances.segments
        keypoints = instances.keypoints
        # Update bboxes if there are segments.
        if len(segments):
            bboxes, segments = self.apply_segments(segments, M)

        if keypoints is not None:
            keypoints = self.apply_keypoints(keypoints, M)
        new_instances = Instances(bboxes, segments, keypoints, bbox_format='xyxy', normalized=False)
        # Clip
        new_instances.clip(*self.size)

        # Filter instances
        instances.scale(scale_w=scale, scale_h=scale, bbox_only=True)
        # Make the bboxes have the same scale with new_bboxes
        i = self.box_candidates(box1=instances.bboxes.T,
                                box2=new_instances.bboxes.T,
                                area_thr=0.01 if len(segments) else 0.10)
        labels['instances'] = new_instances[i]
        labels['cls'] = cls[i]
        labels['img'] = img
        labels['resized_shape'] = img.shape[:2]
        return labels

    def box_candidates(self, box1, box2, wh_thr=2, ar_thr=100, area_thr=0.1, eps=1e-16):  # box1(4,n), box2(4,n)
        # Compute box candidates: box1 before augment, box2 after augment, wh_thr (pixels), aspect_ratio_thr, area_ratio
        w1, h1 = box1[2] - box1[0], box1[3] - box1[1]
        w2, h2 = box2[2] - box2[0], box2[3] - box2[1]
        ar = np.maximum(w2 / (h2 + eps), h2 / (w2 + eps))  # aspect ratio
        return (w2 > wh_thr) & (h2 > wh_thr) & (w2 * h2 / (w1 * h1 + eps) > area_thr) & (ar < ar_thr)  # candidates