二值语义分割标签txt转png格式+语义分割标签筛选
最近在做语义分割路沿检测的任务,需要将多类别的语义分割标签转成0、1值的二值标签,只保留road和background。像cityscapes、
bdd100k、apollo这些数据有.png格式的语义分割标签,还挺好转换的。但是有些标签只有.txt格式,只记录的多边形的顶点坐标,有点像目标检测的标签格式。
接下来就给大家介绍一下语义分割标签如何从txt格式转成png格式,同时还能对标签进行一个筛选,不需要筛选功能的可以自行修改屏蔽掉。
import cv2
import os
import shutil
import json
import numpy as np
images_path = r'C:\seg\0_images/'
labels_path = r'C:\seg\0_labels/'
save_images_path = r'C:\seg\1_images/'
save_labels_path = r'C:\seg\1_labels/'
save_visuals_path = r'C:\seg\1_visual/'
# 提取坐标
def txt2mask_new(img_x, img_y, line):
# 处理每一行的内容
data = line.split('\n')[0]
label = data.split(" ")[0]
d = data.split(' ', -1)
data = []
for i in range(1, int(len(d) / 2) + 1):
data.append([img_y * float(d[2 * i - 1]), img_x * float(d[2 * i])])
data.append(data[0]) # 起点与终点连接
data = np.array(data, dtype=np.int32)
return label, data
#标签选取
def mask_select(img_path,label_path):
#访问txt格式标签文件
with open(label_path, "r") as f1:
new_lines = f1.readlines()
f1.close()
image = cv2.imread(img_path) # 读取图片信息
image_h = image.shape[0]
image_w = image.shape[1]
visual = np.zeros((image_h, image_w, 3), np.uint8) # 生成一张与原图大小相同的值全为0的图
save_label = np.ones((image_h, image_w, 1), np.uint8) # 生成一张与原图大小相同、channels为1的图片
#txt标签处理
for line in new_lines:
label, data = txt2mask_new(image_h, image_w, line)
cv2.fillPoly(visual, [data], color=(0, 0, 255)) #填充多边形,[data]:多边形顶点,围成的区域就是分割区域;color:多边形围成的区域值为(0, 0, 255)
#生成road值为0,background为1的.png格式的语义分割标签
cv2.fillPoly(save_label, [data], color=0) #多边形围成的区域值为0,其余为1。
#图像标签可视化
alpha = 1
beta = 0.5
save_visual = cv2.addWeighted(image, alpha, visual, beta, 0) #融合图像和标签,设置显示比例
cv2.namedWindow('img_save', cv2.WINDOW_NORMAL)
cv2.imshow("img_save", save_visual)
key = cv2.waitKey(0)
#进行可视化筛选,键盘点q退出,点s保存,点其他会到下一张图片。
if key == ord('q'):
cv2.destroyAllWindows()
return 0, image, save_label, save_visual
elif key == ord('s'):
cv2.destroyAllWindows()
return 1, image, save_label, save_visual
else:
cv2.destroyAllWindows()
return 2, image, save_label, save_visual
#查询图片列表
img_items = os.listdir(images_path)
#对图片进行排序,查看自己图片命名自定义设置
# img_items.sort(key=lambda x:( "_".join(x.split("_")[:-1]), int(x.split('_')[-1].split('.jpg')[0])))
# img_items.sort(key=lambda x: (x.split('_')[-3], int(x.split('_')[-2]), int(x.split('_')[-1].split('.jpg')[0])))
# img_items.sort(key=lambda x: ("_".join(x.split("_")[:-2]), int(x.split('_')[-2]), int(x.split('_')[-1].split('.jpg')[0])))
t = 0
for i in img_items:
print(i)
#图像和标签路径
image_path = os.path.join(images_path, i)
label_path = os.path.join(labels_path, i[:-4]+".txt")
save_image_path = save_images_path + '{}.jpg'.format(i[:-4])
save_label_path = save_labels_path + '{}.png'.format(i[:-4])
save_visual_path = save_visuals_path + '{}.jpg'.format(i[:-4])
try:
k, save_image, save_label, save_visual = mask_select(image_path, label_path)
except FileNotFoundError as e:
continue
if k == 0:
break
elif k == 1:
cv2.imwrite(save_image_path, save_image)
cv2.imwrite(save_label_path, save_label)
cv2.imwrite(save_visual_path, save_visual)