DOTAv1数据集格式：
'imagesource':imagesource 'gsd':gsd x1, y1, x2, y2, x3, y3, x4, y4, category, difficult x1, y1, x2, y2, x3, y3, x4, y4, category, difficult ...
imagesource: 图片来源 gsd: 分辨率 x1, y1, x2, y2, x3, y3, x4, y4：四边形的四个顶点的坐标 顶点按顺时针顺序排列，第一个起点为左上第一个点 category:实例类别 difficult：表示该实例是否难以检测（1表示困难，0表示不困难）
COCO转DOTA：
import json
import cv2
import numpy as np
import os
def calculate_rotated_bbox(poly):
"""将多边形坐标转换为旋转边界框"""
contour = np.array(poly).reshape((-1, 1, 2)).astype(np.float32)
rect = cv2.minAreaRect(contour)
box = cv2.boxPoints(rect)
return np.int0(box)
def coco_to_dota(coco_annotation_path, dota_annotation_folder, imagesource="Unknown", gsd="Unknown"):
"""将COCO格式的标注转换为DOTA格式，包括imagesource和gsd信息"""
# 类别ID到名称的映射
category_map = {
1: 'Class1',
2: 'Class2',
}
# 确保输出目录存在
if not os.path.exists(dota_annotation_folder):
os.makedirs(dota_annotation_folder)
# 读取COCO格式的JSON文件
with open(coco_annotation_path, 'r') as f:
coco_data = json.load(f)
# 遍历每个图像的标注
for image in coco_data['images']:
image_id = image['id']
image_filename = image['file_name']
dota_filename = os.path.splitext(image_filename)[0] + '.txt' # 去掉原始扩展名，添加.txt
dota_filepath = os.path.join(dota_annotation_folder, dota_filename)
with open(dota_filepath, 'w') as dota_file:
# 写入imagesource和gsd信息
# dota_file.write(f"'imagesource':{imagesource}\n'gsd':{gsd}\n")
# 找到当前图像的所有标注
for annotation in filter(lambda x: x['image_id'] == image_id, coco_data['annotations']):
if 'segmentation' in annotation:
for seg in annotation['segmentation']:
if type(seg[0]) is list: # 检查是否是多边形格式
seg = seg[0]
box = calculate_rotated_bbox(seg)
# 从映射中获取类别名称
category_name = category_map.get(annotation['category_id'], 'Unknown')
# 格式化DOTA标注
box_str = ' '.join(map(str, box.flatten().tolist()))
dota_annotation = f"{box_str} {category_name} 0\n"
dota_file.write(dota_annotation)
# 调用函数，转换COCO到DOTA
coco_annotation_path = 'instances.json'
dota_annotation_folder = 'dota'
coco_to_dota(coco_annotation_path, dota_annotation_folder)
标注可视化：
import cv2
import numpy as np
import os
def draw_rotated_box(img, box, label):
"""在图像上绘制旋转的边界框和标签。"""
points = np.int0(box)
cv2.drawContours(img, [points], 0, (0, 255, 0), 2) # 绘制旋转框
cv2.putText(img, label, tuple(points[0]), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 255), 1) # 添加文本标签
def visualize_dota_annotations(image_folder, annotation_folder, output_folder):
"""批量处理图像和DOTA标注文件，绘制旋转边界框和标签"""
# 确保输出文件夹存在
if not os.path.exists(output_folder):
os.makedirs(output_folder)
# 遍历图像文件
for img_filename in os.listdir(image_folder):
img_path = os.path.join(image_folder, img_filename)
if os.path.isfile(img_path) and img_filename.endswith(('.jpg', '.png')):
annot_filename = os.path.splitext(img_filename)[0] + '.txt'
annot_path = os.path.join(annotation_folder, annot_filename)
output_img_path = os.path.join(output_folder, img_filename)
# 读取图像
img = cv2.imread(img_path)
if img is None:
continue # 如果图像文件无法读取，则跳过
# 处理对应的标注文件
if os.path.isfile(annot_path):
with open(annot_path, 'r') as f:
lines = f.readlines()
# 跳过文件开头的imagesource和gsd信息
for line in lines[2:]: # 开始处理从第三行起的标注信息
parts = line.strip().split(' ')
if len(parts) < 9: # 跳过非标准行
continue
box = np.array([float(part) for part in parts[:8]]).reshape(4, 2)
label = parts[8] # 标签
draw_rotated_box(img, box, label)
# 保存绘制了旋转边界框的图像
cv2.imwrite(output_img_path, img)
# 路径配置
image_folder = 'images'
annotation_folder = 'dota'
output_folder = 'visual'
visualize_dota_annotations(image_folder, annotation_folder, output_folder)
推荐链接
发表评论