MMDetection 系列之（自定义数据管道处理增强管道）

数据管道设计

遵循典型约定，我们使用Dataset和DataLoader对多个worker进行数据加载。数据集返回与模型的forward方法的参数相对应的数据项字典。由于目标检测中的数据可能不相同大小(图像大小，gt box大小等)，我们在MMCV中引入了一个新的DataContainer类型，以帮助收集和分发不同大小的数据。查看这里了解更多细节。

对数据准备管道和数据集进行分解。数据集通常定义如何处理注释，数据管道定义准备数据字典的所有步骤。管道由一系列操作组成。每个操作都接受一个字典作为输入，并输出一个字典用于下一个转换。
我们在下图中展示了一个经典的管道。蓝色的块是管道操作。随着管道的运行，每个操作符都可以向结果字典添加新的键(标记为绿色)或更新现有的键(标记为橙色)。

1、pipeline example for Faster R-CNN.

img_norm_cfg = dict(
    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations', with_bbox=True),
    dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
    dict(type='RandomFlip', flip_ratio=0.5),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='Pad', size_divisor=32),
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']),
]
test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale=(1333, 800),
        flip=False,
        transforms=[
            dict(type='Resize', keep_ratio=True),
            dict(type='RandomFlip'),
            dict(type='Normalize', **img_norm_cfg),
            dict(type='Pad', size_divisor=32),
            dict(type='ImageToTensor', keys=['img']),
            dict(type='Collect', keys=['img']),
        ])
]

对于每个操作，我们列出了添加/更新/删除的相关dict字段。

预处理

Resize
添加：scale，scale_idx，pad_shape，scale_factor，keep_ratio
更新：img，img_shape，* bbox_fields，* mask_fields，* seg_fields

RandomFlip
添加：翻转
更新：img，* bbox_fields，* mask_fields，* seg_fields

Pad
添加：pad_fixed_size，pad_size_divisor
更新：img，pad_shape，* mask_fields，* seg_fields

RandomCrop
更新：img，pad_shape，gt_bboxes，gt_labels，gt_masks，* bbox_fields

Normalize
添加：img_norm_cfg
更新：img

SegRescale
更新：gt_semantic_seg

PhotoMetricDistortion
更新：img

Expand
更新：img，gt_bboxes

MinIoURandomCrop
更新：img，gt_bboxes，gt_labels

Corrupt
更新：img

格式化
ToTensor

更新：由指定keys。
ImageToTensor

更新：由指定keys。
Transpose

更新：由指定keys。
ToDataContainer

更新：由指定fields。
DefaultFormatBundle

更新：img，投标，gt_bboxes，gt_bboxes_ignore，gt_labels，gt_masks，gt_semantic_seg
Collect

添加：img_meta（img_meta的键由指定meta_keys）
删除：除由所指定的键外的所有其他键 keys
测试时间增加
MultiScaleFlipAug

数据加载

数据加载
LoadImageFromFile
添加：img，img_shape，ori_shape

LoadAnnotations
添加：gt_bboxes，gt_bboxes_ignore，gt_labels，gt_masks，gt_semantic_seg，bbox_fields，mask_fields

LoadProposals
添加：建议

预处理

Resize（缩放）
add: scale, scale_idx, pad_shape, scale_factor, keep_ratio
update: img, img_shape, *bbox_fields, *mask_fields, *seg_fields

RandomFlip（随机镜像）
add: flip
update: img, *bbox_fields, *mask_fields, *seg_fields

Pad（填充）
add: pad_fixed_size, pad_size_divisor
update: img, pad_shape, *mask_fields, *seg_fields

RandomCrop（随机裁剪）
update: img, pad_shape, gt_bboxes, gt_labels, gt_masks, *bbox_fields

Normalize（归一化）
add: img_norm_cfg
update: img

SegRescale
update: gt_semantic_seg

PhotoMetricDistortion
update: img

Expand
update: img, gt_bboxes

MinIoURandomCrop
update: img, gt_bboxes, gt_labels

Corrupt
update: img

格式化

ToTensor
update: specified by keys.

ImageToTensor
update: specified by keys.

Transpose
update: specified by keys.

ToDataContainer
update: specified by fields.

DefaultFormatBundle
update: img, proposals, gt_bboxes, gt_bboxes_ignore, gt_labels, gt_masks, gt_semantic_seg

Collect
add: img_meta (the keys of img_meta is specified by meta_keys)
remove: all other keys except for those specified by keys

扩展和使用自定义管道

1、"在文件中写入一个新的管道，例如my_pipeline.py。它接受一个字典作为输入并返回一个字典。

import random
from mmdet.datasets import PIPELINES


@PIPELINES.register_module()
class MyTransform:
    """Add your transform

    Args:
        p (float): Probability of shifts. Default 0.5.
    """

    def __init__(self, p=0.5):
        self.p = p

    def __call__(self, results):
        if random.random() > self.p:
            results['dummy'] = True
        return results

2. 在配置文件中导入并使用管道。确保导入是相对于train脚本所在的位置。

custom_imports = dict(imports=['path.to.my_pipeline'], allow_failed_imports=False)

img_norm_cfg = dict(
    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations', with_bbox=True),
    dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
    dict(type='RandomFlip', flip_ratio=0.5),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='Pad', size_divisor=32),
    dict(type='MyTransform', p=0.2),
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']),
]

3.可视化您的增强管道的输出

要可视化您的agementpipeline的输出，tools/misc/browse_data .py可以帮助用户可视化地浏览检测数据集(包括图像和边界框注释)，或者将图像保存到指定目录。更多细节可以参考useful_tools

为什么图像和Bbox需要进行数据增强

答：因为数据多了就可以尽可能多的学习到图像中的不变性，学习到的不变性越多那么模型的泛化能力越强。

但是输入到CNN中的图像为什么不具有平移不变性？如何去解决？下面链接有专门的解析：
输入到CNN中的图像为什么不具有平移不变性？如何去解决？

MMdetection 数据增强

MMDetection中，数据增强包括两部分：（源码解析）

新版MMDetection新添加用Albu数据库对图像进行增强的代码，位置在mmdetection/configs/albu_example/mask_rcnn_r50_fpn_1x.py，是基于图像分割的，用于目标检测的代码更新如下

1、引入albumentations数据增强库进行增强

目标检测tricks：Ablu数据库增强

重点部分代码：

albu_train_transforms = [
    # dict(
    #     type='HorizontalFlip',
    #     p=0.5),
    # dict(
    #     type='VerticalFlip',
    #     p=0.5),

    dict(
        type='ShiftScaleRotate',
        shift_limit=0.0625,
        scale_limit=0.0,
        rotate_limit=180,
        interpolation=1,
        p=0.5),
    # dict(
    #     type='RandomBrightnessContrast',
    #     brightness_limit=[0.1, 0.3],
    #     contrast_limit=[0.1, 0.3],
    #     p=0.2),
    # dict(
    #     type='OneOf',
    #     transforms=[
    #         dict(
    #             type='RGBShift',
    #             r_shift_limit=10,
    #             g_shift_limit=10,
    #             b_shift_limit=10,
    #             p=1.0),
    #         dict(
    #             type='HueSaturationValue',
    #             hue_shift_limit=20,
    #             sat_shift_limit=30,
    #             val_shift_limit=20,
    #             p=1.0)
    #     ],
    #     p=0.1),
    # # dict(type='JpegCompression', quality_lower=85, quality_upper=95, p=0.2),
    #
    # dict(type='ChannelShuffle', p=0.1),
    # dict(
    #     type='OneOf',
    #     transforms=[
    #         dict(type='Blur', blur_limit=3, p=1.0),
    #         dict(type='MedianBlur', blur_limit=3, p=1.0)
    #     ],
    #     p=0.1),
]

train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations', with_bbox=True),
    dict(type='Resize', img_scale=[(4096, 800), (4096, 1200)], keep_ratio=True),
    dict(type='RandomFlip', flip_ratio=0.5),

    dict(type='Pad', size_divisor=32),
    '''
    Args:
        transforms (list[dict]): A list of albu transformations , albu transformations 列表
        bbox_params (dict): Bbox_params for albumentation `Compose`
        keymap (dict): Contains {'input key':'albumentation-style key'}
        skip_img_without_anno (bool): Whether to skip the image if no ann left
            after aug 是否跳过图像，如果没有ann离开后aug
    '''
    dict(
        type='Albu',
        transforms=albu_train_transforms,
        bbox_params=dict(
            type='BboxParams',
            format='pascal_voc',
            label_fields=['gt_labels'],
            min_visibility=0.0,
            filter_lost_elements=True),
        keymap={
            'img': 'image',
            'gt_bboxes': 'bboxes'
        },
        update_pad_shape=False,
        skip_img_without_anno=True),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='DefaultFormatBundle'),
    dict(
        type='Collect',
        keys=['img', 'gt_bboxes', 'gt_labels'],
        meta_keys=('filename', 'ori_shape', 'img_shape', 'img_norm_cfg',
                   'pad_shape', 'scale_factor')
    )
]

全部代码

# model settings
model = dict(
    type='CascadeRCNN',
    num_stages=3,
    pretrained='torchvision://resnet50',
    backbone=dict(
        type='ResNet',
        depth=50,
        num_stages=4,
        out_indices=(0, 1, 2, 3),
        frozen_stages=1,
        style='pytorch'),
    neck=dict(
        type='FPN',
        in_channels=[256, 512, 1024, 2048],
        out_channels=256,
        num_outs=5),
    rpn_head=dict(
        type='RPNHead',
        in_channels=256,
        feat_channels=256,
        anchor_scales=[8],
        anchor_ratios=[0.5, 1.0, 2.0],
        anchor_strides=[4, 8, 16, 32, 64],
        target_means=[.0, .0, .0, .0],
        target_stds=[1.0, 1.0, 1.0, 1.0],
        loss_cls=dict(
            type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0), # CrossEntropyLoss/FocalLoss
        loss_bbox=dict(type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=1.0)),
    bbox_roi_extractor=dict(
        type='SingleRoIExtractor',
        roi_layer=dict(type='RoIAlign', out_size=7, sample_num=2),
        out_channels=256,
        featmap_strides=[4, 8, 16, 32]),
    bbox_head=[
        dict(
            type='SharedFCBBoxHead',
            num_fcs=2,
            in_channels=256,
            fc_out_channels=1024,
            roi_feat_size=7,
            num_classes=8,
            target_means=[0., 0., 0., 0.],
            target_stds=[0.1, 0.1, 0.2, 0.2],
            reg_class_agnostic=True,
            loss_cls=dict(
                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
            loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0)),
        dict(
            type='SharedFCBBoxHead',
            num_fcs=2,
            in_channels=256,
            fc_out_channels=1024,
            roi_feat_size=7,
            num_classes=8,
            target_means=[0., 0., 0., 0.],
            target_stds=[0.05, 0.05, 0.1, 0.1],
            reg_class_agnostic=True,
            loss_cls=dict(
                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
            loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0)),
        dict(
            type='SharedFCBBoxHead',
            num_fcs=2,
            in_channels=256,
            fc_out_channels=1024,
            roi_feat_size=7,
            num_classes=8,
            target_means=[0., 0., 0., 0.],
            target_stds=[0.033, 0.033, 0.067, 0.067],
            reg_class_agnostic=True,
            loss_cls=dict(
                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
            loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0))
    ])
# model training and testing settings
train_cfg = dict(
    rpn=dict(
        assigner=dict(
            type='MaxIoUAssigner',
            pos_iou_thr=0.7,
            neg_iou_thr=0.3,
            min_pos_iou=0.3,
            ignore_iof_thr=-1),
        sampler=dict(
            type='RandomSampler',
            num=256,
            pos_fraction=0.5,
            neg_pos_ub=-1,
            add_gt_as_proposals=False),
        allowed_border=0,
        pos_weight=-1,
        debug=False),
    rpn_proposal=dict(
        nms_across_levels=False,
        nms_pre=2000,
        nms_post=2000,
        max_num=2000,
        nms_thr=0.7,
        min_bbox_size=0),
    rcnn=[
        dict(
            assigner=dict(
                type='MaxIoUAssigner',
                pos_iou_thr=0.5,
                neg_iou_thr=0.5,
                min_pos_iou=0.5,
                ignore_iof_thr=-1),
            sampler=dict(
                type='RandomSampler',
                num=512,
                pos_fraction=0.25,
                neg_pos_ub=-1,
                add_gt_as_proposals=True),
            pos_weight=-1,
            debug=False),
        dict(
            assigner=dict(
                type='MaxIoUAssigner',
                pos_iou_thr=0.6,
                neg_iou_thr=0.6,
                min_pos_iou=0.6,
                ignore_iof_thr=-1),
            sampler=dict(
                type='RandomSampler',
                num=512,
                pos_fraction=0.25,
                neg_pos_ub=-1,
                add_gt_as_proposals=True),
            pos_weight=-1,
            debug=False),
        dict(
            assigner=dict(
                type='MaxIoUAssigner',
                pos_iou_thr=0.7,
                neg_iou_thr=0.7,
                min_pos_iou=0.7,
                ignore_iof_thr=-1),
            sampler=dict(
                type='RandomSampler',
                num=512,
                pos_fraction=0.25,
                neg_pos_ub=-1,
                add_gt_as_proposals=True),
            pos_weight=-1,
            debug=False)
    ],
    stage_loss_weights=[1, 0.5, 0.25])
test_cfg = dict(
    rpn=dict(
        nms_across_levels=False,
        nms_pre=1000,
        nms_post=1000,
        max_num=1000,
        nms_thr=0.7,
        min_bbox_size=0),
    rcnn=dict(
        score_thr=0.01,
        nms=dict(type='nms', iou_thr=0.5), max_per_img=100),
    keep_all_stages=False)
# dataset
dataset_type = 'CocoDataset'
data_root = './data/coco/'
img_norm_cfg = dict(
    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)

albu_train_transforms = [
    # dict(
    #     type='HorizontalFlip',
    #     p=0.5),
    # dict(
    #     type='VerticalFlip',
    #     p=0.5),

    dict(
        type='ShiftScaleRotate',
        shift_limit=0.0625,
        scale_limit=0.0,
        rotate_limit=180,
        interpolation=1,
        p=0.5),
    # dict(
    #     type='RandomBrightnessContrast',
    #     brightness_limit=[0.1, 0.3],
    #     contrast_limit=[0.1, 0.3],
    #     p=0.2),
    # dict(
    #     type='OneOf',
    #     transforms=[
    #         dict(
    #             type='RGBShift',
    #             r_shift_limit=10,
    #             g_shift_limit=10,
    #             b_shift_limit=10,
    #             p=1.0),
    #         dict(
    #             type='HueSaturationValue',
    #             hue_shift_limit=20,
    #             sat_shift_limit=30,
    #             val_shift_limit=20,
    #             p=1.0)
    #     ],
    #     p=0.1),
    # # dict(type='JpegCompression', quality_lower=85, quality_upper=95, p=0.2),
    #
    # dict(type='ChannelShuffle', p=0.1),
    # dict(
    #     type='OneOf',
    #     transforms=[
    #         dict(type='Blur', blur_limit=3, p=1.0),
    #         dict(type='MedianBlur', blur_limit=3, p=1.0)
    #     ],
    #     p=0.1),
]

train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations', with_bbox=True),
    dict(type='Resize', img_scale=[(4096, 800), (4096, 1200)], keep_ratio=True),
    dict(type='RandomFlip', flip_ratio=0.5),

    dict(type='Pad', size_divisor=32),
    dict(
        type='Albu',
        transforms=albu_train_transforms,
        bbox_params=dict(
            type='BboxParams',
            format='pascal_voc',
            label_fields=['gt_labels'],
            min_visibility=0.0,
            filter_lost_elements=True),
        keymap={
            'img': 'image',
            'gt_bboxes': 'bboxes'
        },
        update_pad_shape=False,
        skip_img_without_anno=True),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='DefaultFormatBundle'),
    dict(
        type='Collect',
        keys=['img', 'gt_bboxes', 'gt_labels'],
        meta_keys=('filename', 'ori_shape', 'img_shape', 'img_norm_cfg',
                   'pad_shape', 'scale_factor')
    )
]

test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale = (4096, 1000),
        flip=False,
        transforms=[
            dict(type='Resize', keep_ratio=True),
            dict(type='RandomFlip'),
            dict(type='Normalize', **img_norm_cfg),
            dict(type='Pad', size_divisor=32),
            dict(type='ImageToTensor', keys=['img']),
            dict(type='Collect', keys=['img']),
        ])
]
data = dict(
    imgs_per_gpu=2,
    workers_per_gpu=3,
    train=dict(
        type=dataset_type,
        ann_file=data_root + 'first_pinggai/bottle/annotations/instances_train2017.json',
        img_prefix=data_root + 'first_pinggai/bottle/images/train2017/',
        pipeline=train_pipeline),
    val=dict(
        type=dataset_type,
        ann_file=data_root + 'first_pinggai/bottle/annotations/instances_val2017.json',
        img_prefix=data_root + 'first_pinggai/bottle/images/val2017/',
        pipeline=test_pipeline),
    test=dict(
        type=dataset_type,
        ann_file=data_root + 'first_pinggai/bottle/annotations/instances_val2017.json',
        img_prefix=data_root + 'first_pinggai/bottle/images/val2017/',
        pipeline=test_pipeline)
            )
# optimizer
optimizer = dict(type='SGD', lr= 0.005, momentum=0.9, weight_decay=0.0001)
optimizer_config = dict(grad_clip=dict(max_norm=35, norm_type=2))
# learning policy
lr_config = dict(
    policy='step',
    warmup='linear',
    warmup_iters=500,
    warmup_ratio=1.0 / 3,
    step=[8, 11])
checkpoint_config = dict(interval=1)
# yapf:disable
log_config = dict(
    interval=50,
    hooks=[
        dict(type='TextLoggerHook'),
        # dict(type='TensorboardL0oggerHook')
    ])
# yapf:enable
# runtime settings
total_epochs = 50
dist_params = dict(backend='nccl')
log_level = 'INFO'
work_dir = './work_dirs/HRo_cascade50_pinggai_1400'
load_from = './checkpoints/cascade_rcnn_r50_fpn_1x_20190501-3b6211ab.pth'
resume_from = None
workflow = [('train', 1)]

这部分代码需要看开头"注意"部分解析，防止修改代码后运行报错。
需要修改两个部分代码：

mmdet/datasets/pipelines/transforms.py，修改后的源码如下：

transforms.py

import inspect
import albumentations
import mmcv
import numpy as np
from albumentations import Compose
from imagecorruptions import corrupt
from numpy import random

from mmdet.core.evaluation.bbox_overlaps import bbox_overlaps
from ..registry import PIPELINES


@PIPELINES.register_module
class Resize(object):
    """Resize images & bbox & mask.

    This transform resizes the input image to some scale. Bboxes and masks are
    then resized with the same scale factor. If the input dict contains the key
    "scale", then the scale in the input dict is used, otherwise the specified
    scale in the init method is used.

    `img_scale` can either be a tuple (single-scale) or a list of tuple
    (multi-scale). There are 3 multiscale modes:
    - `ratio_range` is not None: randomly sample a ratio from the ratio range
        and multiply it with the image scale.
    - `ratio_range` is None and `multiscale_mode` == "range": randomly sample a
        scale from the a range.
    - `ratio_range` is None and `multiscale_mode` == "value": randomly sample a
        scale from multiple scales.

    Args:
        img_scale (tuple or list[tuple]): Images scales for resizing.
        multiscale_mode (str): Either "range" or "value".
        ratio_range (tuple[float]): (min_ratio, max_ratio)
        keep_ratio (bool): Whether to keep the aspect ratio when resizing the
            image.
    """

    def __init__(self,
                 img_scale=None,
                 multiscale_mode='range',
                 ratio_range=None,
                 keep_ratio=True):
        if img_scale is None:
            self.img_scale = None
        else:
            if isinstance(img_scale, list):
                self.img_scale = img_scale
            else:
                self.img_scale = [img_scale]
            assert mmcv.is_list_of(self.img_scale, tuple)

        if ratio_range is not None:
            # mode 1: given a scale and a range of image ratio
            assert len(self.img_scale) == 1
        else:
            # mode 2: given multiple scales or a range of scales
            assert multiscale_mode in ['value', 'range']

        self.multiscale_mode = multiscale_mode
        self.ratio_range = ratio_range
        self.keep_ratio = keep_ratio

    @staticmethod
    def random_select(img_scales):
        assert mmcv.is_list_of(img_scales, tuple)
        scale_idx = np.random.randint(len(img_scales))
        img_scale = img_scales[scale_idx]
        return img_scale, scale_idx

    @staticmethod
    def random_sample(img_scales):
        assert mmcv.is_list_of(img_scales, tuple) and len(img_scales) == 2
        img_scale_long = [max(s) for s in img_scales]
        img_scale_short = [min(s) for s in img_scales]
        long_edge = np.random.randint(
            min(img_scale_long),
            max(img_scale_long) + 1)
        short_edge = np.random.randint(
            min(img_scale_short),
            max(img_scale_short) + 1)
        img_scale = (long_edge, short_edge)
        return img_scale, None

    @staticmethod
    def random_sample_ratio(img_scale, ratio_range):
        assert isinstance(img_scale, tuple) and len(img_scale) == 2
        min_ratio, max_ratio = ratio_range
        assert min_ratio <= max_ratio
        ratio = np.random.random_sample() * (max_ratio - min_ratio) + min_ratio
        scale = int(img_scale[0] * ratio), int(img_scale[1] * ratio)
        return scale, None

    def _random_scale(self, results):
        if self.ratio_range is not None:
            scale, scale_idx = self.random_sample_ratio(
                self.img_scale[0], self.ratio_range)
        elif len(self.img_scale) == 1:
            scale, scale_idx = self.img_scale[0], 0
        elif self.multiscale_mode == 'range':
            scale, scale_idx = self.random_sample(self.img_scale)
        elif self.multiscale_mode == 'value':
            scale, scale_idx = self.random_select(self.img_scale)
        else:
            raise NotImplementedError

        results['scale'] = scale
        results['scale_idx'] = scale_idx

    def _resize_img(self, results):
        if results['concat']:
            img_raw, img_temp = results['img'][:, :, :3], results['img'][:, :, 3:]
            if self.keep_ratio:
                img_raw, scale_factor = mmcv.imrescale(
                    img_raw, results['scale'], return_scale=True)
                img_temp, scale_factor = mmcv.imrescale(
                    img_temp, results['scale'], return_scale=True)
            else:
                img_raw, w_scale, h_scale = mmcv.imresize(
                    img_raw, results['scale'], return_scale=True)
                img_temp, w_scale, h_scale = mmcv.imresize(
                    img_temp, results['scale'], return_scale=True)
                scale_factor = np.array([w_scale, h_scale, w_scale, h_scale],
                                        dtype=np.float32)
            results['img'] = np.concatenate([img_raw, img_temp], axis=2)
            results['img_shape'] = img_raw.shape
            results['pad_shape'] = img_raw.shape  # in case that there is no padding
            results['scale_factor'] = scale_factor
            results['keep_ratio'] = self.keep_ratio
        else:
            if self.keep_ratio:
                img, scale_factor = mmcv.imrescale(
                    results['img'], results['scale'], return_scale=True)
            else:
                img, w_scale, h_scale = mmcv.imresize(
                    results['img'], results['scale'], return_scale=True)
                scale_factor = np.array([w_scale, h_scale, w_scale, h_scale],
                                        dtype=np.float32)
            results['img'] = img
            results['img_shape'] = img.shape
            results['pad_shape'] = img.shape  # in case that there is no padding
            results['scale_factor'] = scale_factor
            results['keep_ratio'] = self.keep_ratio

    def _resize_bboxes(self, results):
        img_shape = results['img_shape']
        for key in results.get('bbox_fields', []):
            bboxes = results[key] * results['scale_factor']
            bboxes[:, 0::2] = np.clip(bboxes[:, 0::2], 0, img_shape[1] - 1)
            bboxes[:, 1::2] = np.clip(bboxes[:, 1::2], 0, img_shape[0] - 1)
            results[key] = bboxes

    def _resize_masks(self, results):
        for key in results.get('mask_fields', []):
            if results[key] is None:
                continue
            if self.keep_ratio:
                masks = [
                    mmcv.imrescale(
                        mask, results['scale_factor'], interpolation='nearest')
                    for mask in results[key]
                ]
            else:
                mask_size = (results['img_shape'][1], results['img_shape'][0])
                masks = [
                    mmcv.imresize(mask, mask_size, interpolation='nearest')
                    for mask in results[key]
                ]
            results[key] = masks

    def _resize_seg(self, results):
        for key in results.get('seg_fields', []):
            if self.keep_ratio:
                gt_seg = mmcv.imrescale(
                    results[key], results['scale'], interpolation='nearest')
            else:
                gt_seg = mmcv.imresize(
                    results[key], results['scale'], interpolation='nearest')
            results['gt_semantic_seg'] = gt_seg

    def __call__(self, results):
        if 'scale' not in results:
            self._random_scale(results)
        self._resize_img(results)
        self._resize_bboxes(results)
        self._resize_masks(results)
        self._resize_seg(results)
        return results

    def __repr__(self):
        repr_str = self.__class__.__name__
        repr_str += ('(img_scale={}, multiscale_mode={}, ratio_range={}, '
                     'keep_ratio={})').format(self.img_scale,
                                              self.multiscale_mode,
                                              self.ratio_range,
                                              self.keep_ratio)
        return repr_str


@PIPELINES.register_module
class RandomFlip(object):
    """Flip the image & bbox & mask.

    If the input dict contains the key "flip", then the flag will be used,
    otherwise it will be randomly decided by a ratio specified in the init
    method.

    Args:
        flip_ratio (float, optional): The flipping probability.
    """

    def __init__(self, flip_ratio=None, direction='horizontal'):
        self.flip_ratio = flip_ratio
        self.direction = direction
        if flip_ratio is not None:
            assert flip_ratio >= 0 and flip_ratio <= 1
        assert direction in ['horizontal', 'vertical']

    def bbox_flip(self, bboxes, img_shape, direction):
        """Flip bboxes horizontally.

        Args:
            bboxes(ndarray): shape (..., 4*k)
            img_shape(tuple): (height, width)
        """
        assert bboxes.shape[-1] % 4 == 0
        flipped = bboxes.copy()
        if direction == 'horizontal':
            w = img_shape[1]
            flipped[..., 0::4] = w - bboxes[..., 2::4] - 1
            flipped[..., 2::4] = w - bboxes[..., 0::4] - 1
        elif direction == 'vertical':
            h = img_shape[0]
            flipped[..., 1::4] = h - bboxes[..., 3::4] - 1
            flipped[..., 3::4] = h - bboxes[..., 1::4] - 1
        else:
            raise ValueError(
                'Invalid flipping direction "{}"'.format(direction))
        return flipped

    def __call__(self, results):
        if 'flip' not in results:
            flip = True if np.random.rand() < self.flip_ratio else False
            results['flip'] = flip
        if 'flip_direction' not in results:
            results['flip_direction'] = self.direction
        if results['flip']:
            # flip image
            results['img'] = mmcv.imflip(
                results['img'], direction=results['flip_direction'])
            # flip bboxes
            for key in results.get('bbox_fields', []):
                results[key] = self.bbox_flip(results[key],
                                              results['img_shape'],
                                              results['flip_direction'])
            # flip masks
            for key in results.get('mask_fields', []):
                results[key] = [
                    mmcv.imflip(mask, direction=results['flip_direction'])
                    for mask in results[key]
                ]

            # flip segs
            for key in results.get('seg_fields', []):
                results[key] = mmcv.imflip(
                    results[key], direction=results['flip_direction'])
        return results

    def __repr__(self):
        return self.__class__.__name__ + '(flip_ratio={})'.format(
            self.flip_ratio)


@PIPELINES.register_module
class Pad(object):
    """Pad the image & mask.

    There are two padding modes: (1) pad to a fixed size and (2) pad to the
    minimum size that is divisible by some number.

    Args:
        size (tuple, optional): Fixed padding size.
        size_divisor (int, optional): The divisor of padded size.
        pad_val (float, optional): Padding value, 0 by default.
    """

    def __init__(self, size=None, size_divisor=None, pad_val=0):
        self.size = size
        self.size_divisor = size_divisor
        self.pad_val = pad_val
        # only one of size and size_divisor should be valid
        assert size is not None or size_divisor is not None
        assert size is None or size_divisor is None

    def _pad_img(self, results):
        if results['concat']:
            img_raw, img_temp = results['img'][:, :, :3], results['img'][:, :, 3:]
            if self.size is not None:
                padded_img_raw = mmcv.impad(img_raw, self.size)
                padded_img_temp = mmcv.impad(img_temp, self.size)
            elif self.size_divisor is not None:
                padded_img_raw = mmcv.impad_to_multiple(
                    img_raw, self.size_divisor, pad_val=self.pad_val)
                padded_img_temp = mmcv.impad_to_multiple(
                    img_temp, self.size_divisor, pad_val=self.pad_val)
            results['img'] = np.concatenate([padded_img_raw, padded_img_temp], axis=2)
            results['pad_shape'] = padded_img_raw.shape
            results['pad_fixed_size'] = self.size
            results['pad_size_divisor'] = self.size_divisor
        else:
            if self.size is not None:
                padded_img = mmcv.impad(results['img'], self.size)
            elif self.size_divisor is not None:
                padded_img = mmcv.impad_to_multiple(
                    results['img'], self.size_divisor, pad_val=self.pad_val)
            results['img'] = padded_img
            results['pad_shape'] = padded_img.shape
            results['pad_fixed_size'] = self.size
            results['pad_size_divisor'] = self.size_divisor

    def _pad_masks(self, results):
        pad_shape = results['pad_shape'][:2]
        for key in results.get('mask_fields', []):
            padded_masks = [
                mmcv.impad(mask, pad_shape, pad_val=self.pad_val)
                for mask in results[key]
            ]
            if padded_masks:
                results[key] = np.stack(padded_masks, axis=0)
            else:
                results[key] = np.empty((0, ) + pad_shape, dtype=np.uint8)

    def _pad_seg(self, results):
        for key in results.get('seg_fields', []):
            results[key] = mmcv.impad(results[key], results['pad_shape'][:2])

    def __call__(self, results):
        self._pad_img(results)
        self._pad_masks(results)
        self._pad_seg(results)
        return results

    def __repr__(self):
        repr_str = self.__class__.__name__
        repr_str += '(size={}, size_divisor={}, pad_val={})'.format(
            self.size, self.size_divisor, self.pad_val)
        return repr_str


@PIPELINES.register_module
class Normalize(object):
    """Normalize the image.

    Args:
        mean (sequence): Mean values of 3 channels.
        std (sequence): Std values of 3 channels.
        to_rgb (bool): Whether to convert the image from BGR to RGB,
            default is true.
    """

    def __init__(self, mean, std, to_rgb=True):
        self.mean = np.array(mean, dtype=np.float32)
        self.std = np.array(std, dtype=np.float32)
        self.to_rgb = to_rgb

    def __call__(self, results):
        if results['concat']:
            img_raw, img_temp = results['img'][:, :, :3], results['img'][:, :, 3:]
            img_raw = mmcv.imnormalize(img_raw, self.mean, self.std,
                                       self.to_rgb)
            img_temp = mmcv.imnormalize(img_temp, self.mean, self.std,
                                        self.to_rgb)
            results['img'] = np.concatenate([img_raw, img_temp], axis=2)
            results['img_norm_cfg'] = dict(
                mean=self.mean, std=self.std, to_rgb=self.to_rgb)
        else:
            results['img'] = mmcv.imnormalize(results['img'], self.mean, self.std,
                                              self.to_rgb)
            results['img_norm_cfg'] = dict(
                mean=self.mean, std=self.std, to_rgb=self.to_rgb)
        return results

    def __repr__(self):
        repr_str = self.__class__.__name__
        repr_str += '(mean={}, std={}, to_rgb={})'.format(
            self.mean, self.std, self.to_rgb)
        return repr_str


@PIPELINES.register_module
class RandomCrop(object):
    """Random crop the image & bboxes & masks.

    Args:
        crop_size (tuple): Expected size after cropping, (h, w).
    """

    def __init__(self, crop_size):
        self.crop_size = crop_size

    def __call__(self, results):
        img = results['img']
        margin_h = max(img.shape[0] - self.crop_size[0], 0)
        margin_w = max(img.shape[1] - self.crop_size[1], 0)
        offset_h = np.random.randint(0, margin_h + 1)
        offset_w = np.random.randint(0, margin_w + 1)
        crop_y1, crop_y2 = offset_h, offset_h + self.crop_size[0]
        crop_x1, crop_x2 = offset_w, offset_w + self.crop_size[1]

        # crop the image
        img = img[crop_y1:crop_y2, crop_x1:crop_x2, :]
        img_shape = img.shape
        results['img'] = img
        results['img_shape'] = img_shape

        # crop bboxes accordingly and clip to the image boundary
        for key in results.get('bbox_fields', []):
            bbox_offset = np.array([offset_w, offset_h, offset_w, offset_h],
                                   dtype=np.float32)
            bboxes = results[key] - bbox_offset
            bboxes[:, 0::2] = np.clip(bboxes[:, 0::2], 0, img_shape[1] - 1)
            bboxes[:, 1::2] = np.clip(bboxes[:, 1::2], 0, img_shape[0] - 1)
            results[key] = bboxes

        # crop semantic seg
        for key in results.get('seg_fields', []):
            results[key] = results[key][crop_y1:crop_y2, crop_x1:crop_x2]

        # filter out the gt bboxes that are completely cropped
        if 'gt_bboxes' in results:
            gt_bboxes = results['gt_bboxes']
            valid_inds = (gt_bboxes[:, 2] > gt_bboxes[:, 0]) & (
                gt_bboxes[:, 3] > gt_bboxes[:, 1])
            # if no gt bbox remains after cropping, just skip this image
            if not np.any(valid_inds):
                return None
            results['gt_bboxes'] = gt_bboxes[valid_inds, :]
            if 'gt_labels' in results:
                results['gt_labels'] = results['gt_labels'][valid_inds]

            # filter and crop the masks
            if 'gt_masks' in results:
                valid_gt_masks = []
                for i in np.where(valid_inds)[0]:
                    gt_mask = results['gt_masks'][i][crop_y1:crop_y2,
                                                     crop_x1:crop_x2]
                    valid_gt_masks.append(gt_mask)
                results['gt_masks'] = valid_gt_masks

        return results

    def __repr__(self):
        return self.__class__.__name__ + '(crop_size={})'.format(
            self.crop_size)


@PIPELINES.register_module
class SegRescale(object):
    """Rescale semantic segmentation maps.

    Args:
        scale_factor (float): The scale factor of the final output.
    """

    def __init__(self, scale_factor=1):
        self.scale_factor = scale_factor

    def __call__(self, results):
        for key in results.get('seg_fields', []):
            if self.scale_factor != 1:
                results[key] = mmcv.imrescale(
                    results[key], self.scale_factor, interpolation='nearest')
        return results

    def __repr__(self):
        return self.__class__.__name__ + '(scale_factor={})'.format(
            self.scale_factor)


@PIPELINES.register_module
class PhotoMetricDistortion(object):
    """Apply photometric distortion to image sequentially, every transformation
    is applied with a probability of 0.5. The position of random contrast is in
    second or second to last.

    1. random brightness
    2. random contrast (mode 0)
    3. convert color from BGR to HSV
    4. random saturation
    5. random hue
    6. convert color from HSV to BGR
    7. random contrast (mode 1)
    8. randomly swap channels

    Args:
        brightness_delta (int): delta of brightness.
        contrast_range (tuple): range of contrast.
        saturation_range (tuple): range of saturation.
        hue_delta (int): delta of hue.
    """

    def __init__(self,
                 brightness_delta=32,
                 contrast_range=(0.5, 1.5),
                 saturation_range=(0.5, 1.5),
                 hue_delta=18):
        self.brightness_delta = brightness_delta
        self.contrast_lower, self.contrast_upper = contrast_range
        self.saturation_lower, self.saturation_upper = saturation_range
        self.hue_delta = hue_delta

    def __call__(self, results):
        img = results['img']
        # random brightness
        if random.randint(2):
            delta = random.uniform(-self.brightness_delta,
                                   self.brightness_delta)
            img += delta

        # mode == 0 --> do random contrast first
        # mode == 1 --> do random contrast last
        mode = random.randint(2)
        if mode == 1:
            if random.randint(2):
                alpha = random.uniform(self.contrast_lower,
                                       self.contrast_upper)
                img *= alpha

        # convert color from BGR to HSV
        img = mmcv.bgr2hsv(img)

        # random saturation
        if random.randint(2):
            img[..., 1] *= random.uniform(self.saturation_lower,
                                          self.saturation_upper)

        # random hue
        if random.randint(2):
            img[..., 0] += random.uniform(-self.hue_delta, self.hue_delta)
            img[..., 0][img[..., 0] > 360] -= 360
            img[..., 0][img[..., 0] < 0] += 360

        # convert color from HSV to BGR
        img = mmcv.hsv2bgr(img)

        # random contrast
        if mode == 0:
            if random.randint(2):
                alpha = random.uniform(self.contrast_lower,
                                       self.contrast_upper)
                img *= alpha

        # randomly swap channels
        if random.randint(2):
            img = img[..., random.permutation(3)]

        results['img'] = img
        return results

    def __repr__(self):
        repr_str = self.__class__.__name__
        repr_str += ('(brightness_delta={}, contrast_range={}, '
                     'saturation_range={}, hue_delta={})').format(
                         self.brightness_delta, self.contrast_range,
                         self.saturation_range, self.hue_delta)
        return repr_str


@PIPELINES.register_module
class Expand(object):
    """Random expand the image & bboxes.

    Randomly place the original image on a canvas of 'ratio' x original image
    size filled with mean values. The ratio is in the range of ratio_range.

    Args:
        mean (tuple): mean value of dataset.
        to_rgb (bool): if need to convert the order of mean to align with RGB.
        ratio_range (tuple): range of expand ratio.
        prob (float): probability of applying this transformation
    """

    def __init__(self,
                 mean=(0, 0, 0),
                 to_rgb=True,
                 ratio_range=(1, 4),
                 seg_ignore_label=None,
                 prob=0.5):
        self.to_rgb = to_rgb
        self.ratio_range = ratio_range
        if to_rgb:
            self.mean = mean[::-1]
        else:
            self.mean = mean
        self.min_ratio, self.max_ratio = ratio_range
        self.seg_ignore_label = seg_ignore_label
        self.prob = prob

    def __call__(self, results):
        if random.uniform(0, 1) > self.prob:
            return results

        img, boxes = [results[k] for k in ('img', 'gt_bboxes')]

        h, w, c = img.shape
        ratio = random.uniform(self.min_ratio, self.max_ratio)
        expand_img = np.full((int(h * ratio), int(w * ratio), c),
                             self.mean).astype(img.dtype)
        left = int(random.uniform(0, w * ratio - w))
        top = int(random.uniform(0, h * ratio - h))
        expand_img[top:top + h, left:left + w] = img
        boxes = boxes + np.tile((left, top), 2).astype(boxes.dtype)

        results['img'] = expand_img
        results['gt_bboxes'] = boxes

        if 'gt_masks' in results:
            expand_gt_masks = []
            for mask in results['gt_masks']:
                expand_mask = np.full((int(h * ratio), int(w * ratio)),
                                      0).astype(mask.dtype)
                expand_mask[top:top + h, left:left + w] = mask
                expand_gt_masks.append(expand_mask)
            results['gt_masks'] = expand_gt_masks

        # not tested
        if 'gt_semantic_seg' in results:
            assert self.seg_ignore_label is not None
            gt_seg = results['gt_semantic_seg']
            expand_gt_seg = np.full((int(h * ratio), int(w * ratio)),
                                    self.seg_ignore_label).astype(gt_seg.dtype)
            expand_gt_seg[top:top + h, left:left + w] = gt_seg
            results['gt_semantic_seg'] = expand_gt_seg
        return results

    def __repr__(self):
        repr_str = self.__class__.__name__
        repr_str += '(mean={}, to_rgb={}, ratio_range={}, ' \
                    'seg_ignore_label={})'.format(
                        self.mean, self.to_rgb, self.ratio_range,
                        self.seg_ignore_label)
        return repr_str


@PIPELINES.register_module
class MinIoURandomCrop(object):
    """Random crop the image & bboxes, the cropped patches have minimum IoU
    requirement with original image & bboxes, the IoU threshold is randomly
    selected from min_ious.

    Args:
        min_ious (tuple): minimum IoU threshold for all intersections with
        bounding boxes
        min_crop_size (float): minimum crop's size (i.e. h,w := a*h, a*w,
        where a >= min_crop_size).
    """

    def __init__(self, min_ious=(0.1, 0.3, 0.5, 0.7, 0.9), min_crop_size=0.3):
        # 1: return ori img
        self.sample_mode = (1, *min_ious, 0)
        self.min_crop_size = min_crop_size

    def __call__(self, results):
        img, boxes, labels = [
            results[k] for k in ('img', 'gt_bboxes', 'gt_labels')
        ]
        h, w, c = img.shape
        while True:
            mode = random.choice(self.sample_mode)
            if mode == 1:
                return results

            min_iou = mode
            for i in range(50):
                new_w = random.uniform(self.min_crop_size * w, w)
                new_h = random.uniform(self.min_crop_size * h, h)

                # h / w in [0.5, 2]
                if new_h / new_w < 0.5 or new_h / new_w > 2:
                    continue

                left = random.uniform(w - new_w)
                top = random.uniform(h - new_h)

                patch = np.array(
                    (int(left), int(top), int(left + new_w), int(top + new_h)))
                overlaps = bbox_overlaps(
                    patch.reshape(-1, 4), boxes.reshape(-1, 4)).reshape(-1)
                if overlaps.min() < min_iou:
                    continue

                # center of boxes should inside the crop img
                center = (boxes[:, :2] + boxes[:, 2:]) / 2
                mask = ((center[:, 0] > patch[0]) * (center[:, 1] > patch[1]) *
                        (center[:, 0] < patch[2]) * (center[:, 1] < patch[3]))
                if not mask.any():
                    continue
                boxes = boxes[mask]
                labels = labels[mask]

                # adjust boxes
                img = img[patch[1]:patch[3], patch[0]:patch[2]]
                boxes[:, 2:] = boxes[:, 2:].clip(max=patch[2:])
                boxes[:, :2] = boxes[:, :2].clip(min=patch[:2])
                boxes -= np.tile(patch[:2], 2)

                results['img'] = img
                results['gt_bboxes'] = boxes
                results['gt_labels'] = labels

                if 'gt_masks' in results:
                    valid_masks = [
                        results['gt_masks'][i] for i in range(len(mask))
                        if mask[i]
                    ]
                    results['gt_masks'] = [
                        gt_mask[patch[1]:patch[3], patch[0]:patch[2]]
                        for gt_mask in valid_masks
                    ]

                # not tested
                if 'gt_semantic_seg' in results:
                    results['gt_semantic_seg'] = results['gt_semantic_seg'][
                        patch[1]:patch[3], patch[0]:patch[2]]
                return results

    def __repr__(self):
        repr_str = self.__class__.__name__
        repr_str += '(min_ious={}, min_crop_size={})'.format(
            self.min_ious, self.min_crop_size)
        return repr_str


@PIPELINES.register_module
class Corrupt(object):

    def __init__(self, corruption, severity=1):
        self.corruption = corruption
        self.severity = severity

    def __call__(self, results):
        results['img'] = corrupt(
            results['img'].astype(np.uint8),
            corruption_name=self.corruption,
            severity=self.severity)
        return results

    def __repr__(self):
        repr_str = self.__class__.__name__
        repr_str += '(corruption={}, severity={})'.format(
            self.corruption, self.severity)
        return repr_str


@PIPELINES.register_module
class Albu(object):

    def __init__(self,
                 transforms,
                 bbox_params=None,
                 keymap=None,
                 update_pad_shape=False,
                 skip_img_without_anno=False):
        """
        Adds custom transformations from Albumentations lib.
        Please, visit `https://albumentations.readthedocs.io`
        to get more information.

        transforms (list): list of albu transformations
        bbox_params (dict): bbox_params for albumentation `Compose`
        keymap (dict): contains {'input key':'albumentation-style key'}
        skip_img_without_anno (bool): whether to skip the image
                                      if no ann left after aug
        """

        self.transforms = transforms
        self.filter_lost_elements = False
        self.update_pad_shape = update_pad_shape
        self.skip_img_without_anno = skip_img_without_anno

        # A simple workaround to remove masks without boxes
        if (isinstance(bbox_params, dict) and 'label_fields' in bbox_params
                and 'filter_lost_elements' in bbox_params):
            self.filter_lost_elements = True
            self.origin_label_fields = bbox_params['label_fields']
            bbox_params['label_fields'] = ['idx_mapper']
            del bbox_params['filter_lost_elements']

        self.bbox_params = (
            self.albu_builder(bbox_params) if bbox_params else None)
        self.aug = Compose([self.albu_builder(t) for t in self.transforms],
                           bbox_params=self.bbox_params)

        if not keymap:
            self.keymap_to_albu = {
                'img': 'image',
                'gt_masks': 'masks',
                'gt_bboxes': 'bboxes'
            }
        else:
            self.keymap_to_albu = keymap
        self.keymap_back = {v: k for k, v in self.keymap_to_albu.items()}

    def albu_builder(self, cfg):
        """Import a module from albumentations.
        Inherits some of `build_from_cfg` logic.

        Args:
            cfg (dict): Config dict. It should at least contain the key "type".
        Returns:
            obj: The constructed object.
        """
        assert isinstance(cfg, dict) and "type" in cfg
        args = cfg.copy()

        obj_type = args.pop("type")
        if mmcv.is_str(obj_type):
            obj_cls = getattr(albumentations, obj_type)
        elif inspect.isclass(obj_type):
            obj_cls = obj_type
        else:
            raise TypeError(
                'type must be a str or valid type, but got {}'.format(
                    type(obj_type)))

        if 'transforms' in args:
            args['transforms'] = [
                self.albu_builder(transform)
                for transform in args['transforms']
            ]

        return obj_cls(**args)

    @staticmethod
    def mapper(d, keymap):
        """
        Dictionary mapper.
        Renames keys according to keymap provided.

        Args:
            d (dict): old dict
            keymap (dict): {'old_key':'new_key'}
        Returns:
            dict: new dict.
        """
        updated_dict = {}
        for k, v in zip(d.keys(), d.values()):
            new_k = keymap.get(k, k)
            updated_dict[new_k] = d[k]
        return updated_dict

    def __call__(self, results):
        # dict to albumentations format
        results = self.mapper(results, self.keymap_to_albu)

        if 'bboxes' in results:
            # to list of boxes
            if isinstance(results['bboxes'], np.ndarray):
                results['bboxes'] = [x for x in results['bboxes']]
            # add pseudo-field for filtration
            if self.filter_lost_elements:
                results['idx_mapper'] = np.arange(len(results['bboxes']))

        results = self.aug(**results)

        if 'bboxes' in results:
            if isinstance(results['bboxes'], list):
                results['bboxes'] = np.array(
                    results['bboxes'], dtype=np.float32)

            # filter label_fields
            if self.filter_lost_elements:

                results['idx_mapper'] = np.arange(len(results['bboxes']))

                for label in self.origin_label_fields:
                    results[label] = np.array(
                        [results[label][i] for i in results['idx_mapper']])
                if 'masks' in results:
                    results['masks'] = [
                        results['masks'][i] for i in results['idx_mapper']
                    ]

                if (not len(results['idx_mapper'])
                        and self.skip_img_without_anno):
                    return None

        if 'gt_labels' in results:
            if isinstance(results['gt_labels'], list):
                results['gt_labels'] = np.array(results['gt_labels'])

        # back to the original format
        results = self.mapper(results, self.keymap_back)

        # update final shape
        if self.update_pad_shape:
            results['pad_shape'] = results['img'].shape

        return results

    def __repr__(self):
        repr_str = self.__class__.__name__
        repr_str += '(transformations={})'.format(self.transformations)
        return repr_str

mmdet/datasets/pipelines/init.py，修改后的源码如下：

init.py

此代码为引入albumentations数据增强库进行增强

具体修改方式如下：添加dict(type=‘Albu’, transforms = [{“type”: ‘RandomRotate90’}])，其他的类似。

from .compose import Compose
from .formating import (Collect, ImageToTensor, ToDataContainer, ToTensor,
                        Transpose, to_tensor)
from .loading import LoadAnnotations, LoadImageFromFile, LoadProposals
from .test_aug import MultiScaleFlipAug
from .transforms import (Albu, Expand, MinIoURandomCrop, Normalize, Pad,
                         PhotoMetricDistortion, RandomCrop, RandomFlip, Resize,
                         SegRescale)


__all__ = [
    'Compose', 'to_tensor', 'ToTensor', 'ImageToTensor', 'ToDataContainer',
    'Transpose', 'Collect', 'LoadAnnotations', 'LoadImageFromFile',
    'LoadProposals', 'MultiScaleFlipAug', 'Resize', 'RandomFlip', 'Pad',
    'RandomCrop', 'Normalize', 'SegRescale', 'MinIoURandomCrop', 'Expand',
    'PhotoMetricDistortion', 'Albu'
]

具体修改方式如下：添加dict(type=‘Albu’, transforms = [{“type”: ‘RandomRotate90’}])，其他的类似。

train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations', with_bbox=True),
    dict(type='Albu', transforms = [{"type": 'RandomRotate90'}]),# 数据增强
    dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
    dict(type='RandomFlip', flip_ratio=0.5),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='Pad', size_divisor=32),
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']),
]

目前的transform v2

from .auto_augment import (AutoAugment, BrightnessTransform, ColorTransform,
                           ContrastTransform, EqualizeTransform, Rotate, Shear,
                           Translate)
from .compose import Compose
from .formating import (Collect, DefaultFormatBundle, ImageToTensor,
                        ToDataContainer, ToTensor, Transpose, to_tensor)
from .instaboost import InstaBoost
from .loading import (LoadAnnotations, LoadImageFromFile, LoadImageFromWebcam,
                      LoadMultiChannelImageFromFiles, LoadProposals)
from .test_time_aug import MultiScaleFlipAug
from .transforms import (Albu, CutOut, Expand, MinIoURandomCrop, Normalize,
                         Pad, PhotoMetricDistortion, RandomCenterCropPad,
                         RandomCrop, RandomFlip, Resize, SegRescale)

__all__ = [
    'Compose', 'to_tensor', 'ToTensor', 'ImageToTensor', 'ToDataContainer',
    'Transpose', 'Collect', 'DefaultFormatBundle', 'LoadAnnotations',
    'LoadImageFromFile', 'LoadImageFromWebcam',
    'LoadMultiChannelImageFromFiles', 'LoadProposals', 'MultiScaleFlipAug',
    'Resize', 'RandomFlip', 'Pad', 'RandomCrop', 'Normalize', 'SegRescale',
    'MinIoURandomCrop', 'Expand', 'PhotoMetricDistortion', 'Albu',
    'InstaBoost', 'RandomCenterCropPad', 'AutoAugment', 'CutOut', 'Shear',
    'Rotate', 'ColorTransform', 'EqualizeTransform', 'BrightnessTransform',
    'ContrastTransform', 'Translate'
]

2、MMDetection自带数据增强

源码在mmdet/datasets/extra_aug.py里面，包括RandomCrop、brightness、contrast、saturation、ExtraAugmentation等等图像增强方法。
添加config位置是train_pipeline或test_pipeline这个地方（一般train进行增强而test不需要），例如数据增强RandomFlip，flip_ratio代表随机翻转的概率：

train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations', with_bbox=True),
    dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
    dict(type='RandomFlip', flip_ratio=0.5),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='Pad', size_divisor=32),
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']),
]
test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale=(1333, 800),
        flip=False,
        transforms=[
            dict(type='Resize', keep_ratio=True),
            dict(type='RandomFlip'),
            dict(type='Normalize', **img_norm_cfg),
            dict(type='Pad', size_divisor=32),
            dict(type='ImageToTensor', keys=['img']),
            dict(type='Collect', keys=['img']),
        ])
]

3、Bbox增强

albumentations数据增强（同上）
源码在mmdet/datasets/custom.py里面，增强源码为

def pre_pipeline(self, results):
        results['img_prefix'] = self.img_prefix
        results['seg_prefix'] = self.seg_prefix
        results['proposal_file'] = self.proposal_file
        results['bbox_fields'] = []
        results['mask_fields'] = []

配置文件增加的数据增强

configs/base/datasets/coco_detection.py 在train pipeline修改Data Augmentation在train

dataset_type = 'CocoDataset'
data_root = 'data/coco/'
img_norm_cfg = dict(
    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
# 在这里加albumentation的aug
albu_train_transforms = [
    dict(
        type='ShiftScaleRotate',
        shift_limit=0.0625,
        scale_limit=0.0,
        rotate_limit=0,
        interpolation=1,
        p=0.5),
    dict(
        type='RandomBrightnessContrast',
        brightness_limit=[0.1, 0.3],
        contrast_limit=[0.1, 0.3],
        p=0.2),
    dict(
        type='OneOf',
        transforms=[
            dict(
                type='RGBShift',
                r_shift_limit=10,
                g_shift_limit=10,
                b_shift_limit=10,
                p=1.0),
            dict(
                type='HueSaturationValue',
                hue_shift_limit=20,
                sat_shift_limit=30,
                val_shift_limit=20,
                p=1.0)
        ],
        p=0.1),
    dict(type='JpegCompression', quality_lower=85, quality_upper=95, p=0.2),
    dict(type='ChannelShuffle', p=0.1),
    dict(
        type='OneOf',
        transforms=[
            dict(type='Blur', blur_limit=3, p=1.0),
            dict(type='MedianBlur', blur_limit=3, p=1.0)
        ],
        p=0.1),
]
train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations', with_bbox=True, with_mask=True),
    #据说这里改img_scale即可多尺度训练，但是实际运行报错。
    dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
    dict(type='Pad', size_divisor=32),
    dict(
        type='Albu',
        transforms=albu_train_transforms,
        bbox_params=dict(
            type='BboxParams',
            format='pascal_voc',
            label_fields=['gt_labels'],
            min_visibility=0.0,
            filter_lost_elements=True),
        keymap={
            'img': 'image',
            'gt_masks': 'masks',
            'gt_bboxes': 'bboxes'
        },
]
# 测试的pipeline
test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        # 多尺度测试 TTA在这里修改，注意有些模型不支持多尺度TTA，比如cascade_mask_rcnn，若不支持会提示
        # Unimplemented Error
        img_scale=(1333, 800),
        flip=False,
        transforms=[
            dict(type='Resize', keep_ratio=True),
            dict(type='RandomFlip'),
            dict(type='Normalize', **img_norm_cfg),
            dict(type='Pad', size_divisor=32),
            dict(type='ImageToTensor', keys=['img']),
            dict(type='Collect', keys=['img']),
        ])
]
# 包含batch_size, workers和路径。
# 路径如果按照上面的设置好就不需要更改
data = dict(
    samples_per_gpu=2,
    workers_per_gpu=2,
    train=dict(
        type=dataset_type,
        ann_file=data_root + 'annotations/instances_train2017.json',
        img_prefix=data_root + 'train2017/',
        pipeline=train_pipeline),
    val=dict(
        type=dataset_type,
        ann_file=data_root + 'annotations/instances_val2017.json',
        img_prefix=data_root + 'val2017/',
        pipeline=test_pipeline),
    test=dict(
        type=dataset_type,
        ann_file=data_root + 'annotations/instances_val2017.json',
        img_prefix=data_root + 'val2017/',
        pipeline=test_pipeline))
evaluation = dict(interval=1, metric='bbox')