SSD代码解读（二）——Data Augmentation

SSD系列代码解读：（一） Prior Box

SSD系列代码解读：（二） Data Augmentation

SSD系列代码解读：（三） MultiboxLoss

本部分代码是pytorch版本的，非官方的caffe实现，贴上代码解读的同时会与caffe实现进行比较。先贴代码

import torch
from torchvision import transforms
import cv2
import numpy as np
import random
import math
from utils.box_utils import matrix_iou

def _crop(image, boxes, labels):
    height, width, _ = image.shape

    if len(boxes)== 0:
        return image, boxes, labels

    while True: # caffe中的min_iou多了个1.0
        mode = random.choice((
            None,
            (0.1, None),
            (0.3, None),
            (0.5, None),
            (0.7, None),
            (0.9, None),
            (None, None),
        ))

        if mode is None: #随到None，直接返回，1/6概率
            return image, boxes, labels

        min_iou, max_iou = mode
        if min_iou is None:
            min_iou = float('-inf')
        if max_iou is None:
            max_iou = float('inf')

        for _ in range(50): #最大重复裁剪50次，直到某次裁剪合格
            # 面积比是scale^2, aspect_ratio是长宽比，从而获得img_n的w和h
            scale = random.uniform(0.3,1.)
            min_ratio = max(0.5, scale*scale)
            max_ratio = min(2, 1. / scale / scale)
            ratio = math.sqrt(random.uniform(min_ratio, max_ratio))
            w = int(scale * ratio * width)
            h = int((scale / ratio) * height)

            # 随机生成img_n的左上角点坐标，进而获得img_n的位置，就是roi
            l = random.randrange(width - w)
            t = random.randrange(height - h)
            roi = np.array((l, t, l + w, t + h))

            iou = matrix_iou(boxes, roi[np.newaxis])
            
            # 若不存在任何一个GT与roi的iou大于之前随机的iou_min，则重新裁剪
            if not (min_iou <= iou.min() and iou.max() <= max_iou):
                continue
            
            # 获取img_n的像素信息，注意height是第一维
            image_t = image[roi[1]:roi[3], roi[0]:roi[2]]
            
            # 仅保留GT中心在img_n的img_n，若没有，则重新裁剪
            centers = (boxes[:, :2] + boxes[:, 2:]) / 2
            mask = np.logical_and(roi[:2] < centers, centers < roi[2:]) \
                     .all(axis=1)
            boxes_t = boxes[mask].copy()
            labels_t = labels[mask].copy()
            if len(boxes_t) == 0:
                continue
            # 对GT的坐标重新限定，主要是因为边界问题
            boxes_t[:, :2] = np.maximum(boxes_t[:, :2], roi[:2])
            boxes_t[:, :2] -= roi[:2]
            boxes_t[:, 2:] = np.minimum(boxes_t[:, 2:], roi[2:])
            boxes_t[:, 2:] -= roi[:2]
            # 返回裁剪后的img，box和label信息
            return image_t, boxes_t,labels_t

# 亮度对比度在RGB空间调整，色相饱和度在HSV空间调整，都是以0.5的概率
def _distort(image):
    def _convert(image, alpha=1, beta=0):
        tmp = image.astype(float) * alpha + beta
        tmp[tmp < 0] = 0
        tmp[tmp > 255] = 255
        image[:] = tmp

    image = image.copy()

    if random.randrange(2):
        _convert(image, beta=random.uniform(-32, 32))

    if random.randrange(2):
        _convert(image, alpha=random.uniform(0.5, 1.5))

    image = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)

    if random.randrange(2):
        tmp = image[:, :, 0].astype(int) + random.randint(-18, 18)
        tmp %= 180
        image[:, :, 0] = tmp

    if random.randrange(2):
        _convert(image[:, :, 1], alpha=random.uniform(0.5, 1.5))

    image = cv2.cvtColor(image, cv2.COLOR_HSV2BGR)

    return image


# 扩展图片，以p的概率，caffe中p=0.5，pytorch中p=0.6
def _expand(image, boxes,fill, p):
    if random.random() > p:
        return image, boxes

    height, width, depth = image.shape
    for _ in range(50): # 最大重复实验50次
        scale = random.uniform(1,4)

        min_ratio = max(0.5, 1./scale/scale)
        max_ratio = min(2, scale*scale)
        ratio = math.sqrt(random.uniform(min_ratio, max_ratio))
        ws = scale*ratio
        hs = scale/ratio
        if ws < 1 or hs < 1: # 扩展后的长和宽必须都要大于1
            continue
        w = int(ws * width)
        h = int(hs * height)
        
        # 随机生成左上角的点的坐标
        left = random.randint(0, w - width)
        top = random.randint(0, h - height)
        
        # 对GT的坐标的调整
        boxes_t = boxes.copy()
        boxes_t[:, :2] += (left, top)
        boxes_t[:, 2:] += (left, top)

        # 扩展后的图像，和原图重叠部分原像素填充；其他部分填充均值，因为后续需要减去均值，所以等价于0填充，即为黑边
        expand_image = np.empty(
            (h, w, depth),
            dtype=image.dtype)
        expand_image[:, :] = fill
        expand_image[top:top + height, left:left + width] = image
        image = expand_image

        return image, boxes_t

# 以0.5的概率水平翻转,返回处理后的图片和GT信息
def _mirror(image, boxes):
    _, width, _ = image.shape
    if random.randrange(2):
        image = image[:, ::-1]
        boxes = boxes.copy()
        boxes[:, 0::2] = width - boxes[:, 2::-2]
    return image, boxes

# 随机选择一种resize方式，进行resize，并将channel维度调到第一维
def preproc_for_test(image, insize, mean):
    interp_methods = [cv2.INTER_LINEAR, cv2.INTER_CUBIC, cv2.INTER_AREA, cv2.INTER_NEAREST, cv2.INTER_LANCZOS4]
    interp_method = interp_methods[random.randrange(5)]
    image = cv2.resize(image, (insize, insize),interpolation=interp_method)
    image = image.astype(np.float32)
    image -= mean
    return image.transpose(2, 0, 1)

# 数据增强类
class preproc(object):

    def __init__(self, resize, rgb_means, p):
        self.means = rgb_means
        self.resize = resize
        self.p = p

    def __call__(self, image, targets): 
        # targets.shape = (n,5),n是img中target的数量，5是(x1,y1,x2,y2,label)
        # image 是原图
        boxes = targets[:,:-1].copy()
        labels = targets[:,-1].copy()
        if len(boxes) == 0: # 若img中没有gt, resize后再减去均值直接返回
            #boxes = np.empty((0, 4))
            targets = np.zeros((1,5))
            image = preproc_for_test(image, self.resize, self.means)
            return torch.from_numpy(image), targets
        
        # 下面的代码段实现拷贝作用，备份。
        image_o = image.copy()
        targets_o = targets.copy()
        height_o, width_o, _ = image_o.shape
        boxes_o = targets_o[:,:-1]
        labels_o = targets_o[:,-1]
        boxes_o[:, 0::2] /= width_o
        boxes_o[:, 1::2] /= height_o
        labels_o = np.expand_dims(labels_o,1)
        targets_o = np.hstack((boxes_o,labels_o))
        
        #数据增强部分
        image_t, boxes, labels = _crop(image, boxes, labels) # 先裁剪
        image_t = _distort(image_t) # 亮度对比度色相饱和度等属性调整
        image_t, boxes = _expand(image_t, boxes, self.means, self.p) #裁剪后再扩展
        image_t, boxes = _mirror(image_t, boxes) # 水平翻转

        height, width, _ = image_t.shape
        image_t = preproc_for_test(image_t, self.resize, self.means) # aug后的img进行resize并减去均值
        # GT信息不参与resize操作，并将其转化为和resize前aug后的图片的百分比形式，因为SSD的预测信息是小数。
        boxes = boxes.copy()
        boxes[:, 0::2] /= width
        boxes[:, 1::2] /= height
        b_w = (boxes[:, 2] - boxes[:, 0])*1.
        b_h = (boxes[:, 3] - boxes[:, 1])*1.
        mask_b= np.minimum(b_w, b_h) > 0.01 # 太小的GT排除掉
        boxes_t = boxes[mask_b]
        labels_t = labels[mask_b].copy()

        if len(boxes_t)==0: #若aug后的img都是太小的GT，则取消aug，直接对原图resize并剪均值。
            image = preproc_for_test(image_o, self.resize, self.means) #此处体现了之前备份的作用
            return torch.from_numpy(image),targets_o

        labels_t = np.expand_dims(labels_t,1)
        targets_t = np.hstack((boxes_t,labels_t)) #整合targets信息

        return torch.from_numpy(image_t), targets_t



class BaseTransform(object):
    """Defines the transformations that should be applied to test PIL image
        for input into the network
    dimension -> tensorize -> color adj
    Arguments:
        resize (int): input dimension to SSD
        rgb_means ((int,int,int)): average RGB of the dataset
            (104,117,123)
        swap ((int,int,int)): final order of channels
    Returns:
        transform (transform) : callable transform to be applied to test/val
        data
    """
    def __init__(self, resize, rgb_means, swap=(2, 0, 1)):
        self.means = rgb_means
        self.resize = resize
        self.swap = swap

    # assume input is cv2 img for now
    def __call__(self, img):

        interp_methods = [cv2.INTER_LINEAR, cv2.INTER_CUBIC, cv2.INTER_AREA, cv2.INTER_NEAREST, cv2.INTER_LANCZOS4]
        interp_method = interp_methods[0]
        img = cv2.resize(np.array(img), (self.resize,
                                         self.resize),interpolation = interp_method).astype(np.float32)
        img -= self.means
        img = img.transpose(self.swap)
        return torch.from_numpy(img)

整个Aug的流程是： 

crop的操作最为复杂：

1. 首先随机选取iou_min;

2. 随机选择scale(<1)，进行面积上的缩放，再随机出aspect_ratio， 进行长宽比缩放，进一步随机出crop区域的左上角坐标，从而确定crop区域；

3. 判断是否存在一个GT与crop区域的 iou > iou_min，若不存在，最大重复实验50次；

4. 满足3后，再筛选出满足GT的中心点在crop区域的crop区域，若没有，最大重复实验50次；

5，针对crop区域，修改GT信息，主要是边界的调整。

 

expand的操作如下：

1. 以一定的概率进行扩展；

2. 随机生成scale(>1)和aspect_ratio, 判断扩展后的width和height是否都大于未扩展前的，若不满足，最大重复实验50次；

3. 随机生成扩展图像的左上角坐标，并修改GT信息，主要是进行一个平移；

4. 扩展后的图像像素填充。未扩展前图像的那部分原像素填充，其他部分则均值填充，因为后续还需要减去均值，所以等价于0值填充。