Nanodet自带数据增强

1.自己写的数据增强

from torchvision import transforms
from albumentations import (
    HorizontalFlip, IAAPerspective, ShiftScaleRotate, CLAHE, RandomRotate90,
    Transpose, ShiftScaleRotate, Blur,GaussianBlur, OpticalDistortion, GridDistortion, HueSaturationValue,
    IAAAdditiveGaussianNoise, GaussNoise, MotionBlur, MedianBlur, IAAPiecewiseAffine,
    IAASharpen, IAAEmboss, RandomBrightnessContrast, Flip, OneOf, Compose,RandomBrightness,ToSepia, BboxParams,
    VerticalFlip,RandomScale,RandomSizedBBoxSafeCrop,Rotate
)
import albumentations as A

import numpy as np
# 0.5
def strong_aug(p=0.9):
    return Compose([
        OneOf([
            IAAAdditiveGaussianNoise(), #�~F�~X�~V��~Y�声添�~J| �~H��~S�~E��~[��~C~O
            GaussNoise(), # �~F�~X�~V��~Y�声�~T�~T��~N�~S�~E��~[��~C~O
        ], p=0.2),
        OneOf([
            MotionBlur(p=0.25), #使�~T��~Z~O�~\�大�~O�~Z~D�~F~E�| ��~F�~P�~J�模�~J�~T�~T��~N�~S�~E��~[��~C~O
            GaussianBlur(p=0.5),
            Blur(blur_limit=3, p=0.25), #使�~T��~Z~O�~\�大�~O�~Z~D�~F~E�| �模�~J�~S�~E��~[��~C~O
        ], p=0.2),
        HueSaturationValue(p=0.2),
        OneOf([
            RandomBrightness(),
            IAASharpen(),
            IAAEmboss(),
            RandomBrightnessContrast(),
        ], p=0.6),
        ToSepia(p=0.1) #�~U�~P�~I�
    ], p=p)

# 0.7
def aug_with_bbox(p=0.7):
    return Compose([
        HorizontalFlip(p=0.6), #水平�
        VerticalFlip(p=0.6),
        RandomScale(scale_limit=0.1, interpolation=1, always_apply=False, p=0.5),
        RandomSizedBBoxSafeCrop(width=320, height=320, erosion_rate=0.2, p=0.5),
        Rotate(limit=90, p=0.5),
        # �~Z~O�~\�仿�~D�~O~X�~M�
        ShiftScaleRotate(shift_limit=0.0625, scale_limit=0.50, rotate_limit=45, p=.75)],
        p=p,
        bbox_params=BboxParams(format='pascal_voc', min_area=1024, min_visibility=0.1, label_fields=['class_labels']))

def aug_all(p=0.7):
    return Compose([
        OneOf([
            IAAAdditiveGaussianNoise(), #�~F�~X�~V��~Y�声添�~J| �~H��~S�~E��~[��~C~O
            GaussNoise(), # �~F�~X�~V��~Y�声�~T�~T��~N�~S�~E��~[��~C~O
        ], p=0.2),
        OneOf([
            MotionBlur(p=0.25), #使�~T��~Z~O�~\�大�~O�~Z~D�~F~E�| ��~F�~P�~J�模�~J�~T�~T��~N�~S�~E��~[��~C~O
            GaussianBlur(p=0.5),
            Blur(blur_limit=3, p=0.25), #使�~T��~Z~O�~\�大�~O�~Z~D�~F~E�| �模�~J�~S�~E��~[��~C~O
        ], p=0.2),
        HueSaturationValue(p=0.2),
        OneOf([
            RandomBrightness(),
            IAASharpen(),
            IAAEmboss(),
            RandomBrightnessContrast(),
        ], p=0.6),
        ToSepia(p=0.1), #�~U�~P�
        HorizontalFlip(p=0.6), #水平�
        VerticalFlip(p=0.6),
        RandomScale(scale_limit=0.1, interpolation=1, always_apply=False, p=0.5),
        RandomSizedBBoxSafeCrop(width=320, height=320, erosion_rate=0.2, p=0.5),
        Rotate(limit=90, p=0.5),
        # �~Z~O�~\�仿�~D�~O~X�~M�
        ShiftScaleRotate(shift_limit=0.0625, scale_limit=0.50, rotate_limit=90, p=0.5)
    ], p=p, bbox_params=BboxParams(format='pascal_voc', min_area=1025, min_visibility=0.1, label_fields=['class_labels']))

很多中文乱了，其实第3个函数是对前两个的合并，第1个函数包括了高斯噪声之类的非物理位置变化数据增强；第2个函数是包含翻转之类的物理位置变化的数据增强。

self.tranform = 增强函数
transformed = self.transform(image=meta["img"], bboxes=meta["gt_bboxes"], class_labels=meta["gt_labels"])
            meta["img"] = transformed["image"]
            meta["gt_bboxes"] = np.array(transformed["bboxes"])
            meta["gt_labels"] = np.array(transformed["class_labels"])

使用如上。 

2.nanodet的数据增强

    pipeline:
      perspective: 0.0
      scale: [0.6, 1.4]
      stretch: [[0.8, 1.2], [0.8, 1.2]]
      rotation: 0
      shear: 0
      translate: 0.2
      flip: 0.5
      brightness: 0.2
      contrast: [0.6, 1.4]
      saturation: [0.5, 1.2]
      normalize: [[103.53, 116.28, 123.675], [57.375, 57.12, 58.395]]

2.1 perspective 透视变换

此图参考了 (1条消息) yolov5数据增强引发的思考——透视变换矩阵的创建_之子无裳的博客-CSDN博客

可以看到，透视变换之后，标注框变形不再是矩形框，不便训练，因此一般不使用。 设置值为0就好了。

2.2 scale 缩放

2.3 stretch 拉伸

2.4 rotation 旋转

2.5 shear 切割

2.6 translate 平移

2.7 flip 翻转

2.8 brightness 亮度

2.9 contrast 对比度

2.10 saturation 饱和度