Tensorflow object detection api 使用和数据增强旋转

前面的部分参考TensorFlow 训练自己的数据集（包括修改文件等等）
https://blog.csdn.net/int93/article/details/79064428

项目结构

my_train
    — dataset
    — models
    — object_detection
    — create_pascal_tf_record.py
    — eval.py
    — export_inference_graph.py
    — train.py
    — img_augmentation.py
    — pascal_label_map.pbtxt
    — ssd_mobilenet_v2_coco.config

创建tf record文件

python3 create_pascal_tf_record.py 
--data_dir=/Users/anida.qin/Desktop/Projects/Work_02/obj_train/my_train/dataset
--year=my_VOC 
--set=val
--output_path=/Users/anida.qin/Desktop/Projects/Work_02/obj_train/my_train/val.record 
--label_map_path=/Users/anida.qin/Desktop/Projects/Work_02/obj_train/my_train/pascal_label_map.pbtxt

python3 create_pascal_tf_record.py 
--data_dir=/Users/anida.qin/Desktop/Projects/Work_02/obj_train/my_train/dataset 
--year=my_VOC 
--set=train 
--output_path=/Users/anida.qin/Desktop/Projects/Work_02/obj_train/my_train/train.record 
--label_map_path=/Users/anida.qin/Desktop/Projects/Work_02/obj_train/my_train/pascal_label_map.pbtxt

训练

nohup python3 train.py --logtostderr &

输出.pb模型文件

python export_inference_graph.py 
--input_type image_tensor 
--pipeline_config_path ssd_mobilenet_v2_coco.config 
--trained_checkpoint_prefix train/model.ckpt-200000 --output_directory train/

图片旋转原理见 https://www.oschina.net/translate/opencv-rotation?print
有两种思路

思路1

• 先以图片中心旋转图片，得到旋转矩阵
• 然后平移到新的图片尺寸下（不会被遮挡的）的图像中心
• (Rp+t)+m 所以m=(new_w-old_w)/2, (new_h-old_h)/2
• 代码 https://www.cnblogs.com/darkknightzh/p/5070576.html

思路2

• 以新图片尺寸的中心为轴，得到旋转矩阵
• 然后再计算平移量
• R(Ip+m)+t=Rp+Rm+t 所以应该求Rm, m和思路1 的m一样，但是要乘R
• 代码 https://www.oschina.net/translate/opencv-rotation?print

注意
因为这个旋转会修改图片尺寸，因此千万别忘记修改xml文件里的图片大小。因为bbox regression需要用的，影响loss的。
具体解释见 https://blog.csdn.net/weixin_42280271/article/details/82052614

#!/usr/bin/env python

import cv2
import math
import numpy as np
import os
import xml.etree.ElementTree as ET
import random

class ImgAugemention():
    def __init__(self):
        self.angle = 90

    # rotate_img
    def rotate_image(self, src, angle, scale=1.):
        w = src.shape[1]
        h = src.shape[0]
        # convet angle into rad
        rangle = np.deg2rad(angle)  # angle in radians
        # calculate new image width and height
        nw = (abs(np.sin(rangle)*h) + abs(np.cos(rangle)*w))*scale
        nh = (abs(np.cos(rangle)*h) + abs(np.sin(rangle)*w))*scale
        # ask OpenCV for the rotation matrix
        rot_mat = cv2.getRotationMatrix2D((nw*0.5, nh*0.5), angle, scale)
        # calculate the move from the old center to the new center combined