官方的RTDETR是使用COCO格式的数据集进行训练的,自然是COCO指标,但是当我们使用yolov8代码中的进行RTDETR进行训练时,最终现实的是YOLO的指标,如果用到论文里面可能需要转换一下格式,因此讲一下如何转换指标
1.获得data.json文件
自己的数据集是这样的:
yolo2coco.py
import os
import cv2
import json
from tqdm import tqdm
from sklearn.model_selection import train_test_split
import argparse
# 根据自己的数据集类别进行修改
classes =['car', 'truck', 'bus']
parser = argparse.ArgumentParser()
# 验证集图像的路径
parser.add_argument('--image_path', default='/home/root1/data/miya/yolov8/data/uavdt/val/images',type=str, help="path of images")
# 验证集标签的路径
parser.add_argument('--label_path', default='/home/root1/data/miya/yolov8/data/uavdt/val/labels',type=str, help="path of labels .txt")
# 生成data.json的路径
parser.add_argument('--save_path', type=str,default='data.json', help="if not split the dataset, give a path to a json file")
arg = parser.parse_args()
def yolo2coco(arg):
print("Loading data from ", arg.image_path, arg.label_path)
assert os.path.exists(arg.image_path)
assert os.path.exists(arg.label_path)
originImagesDir = arg.image_path
originLabelsDir = arg.label_path
# images dir name
indexes = os.listdir(originImagesDir)
dataset = {'categories': [], 'annotations': [], 'images': []}
for i, cls in enumerate(classes, 0):
dataset['categories'].append({'id': i, 'name': cls, 'supercategory': 'mark'})
# 标注的id
ann_id_cnt = 0
for k, index in enumerate(tqdm(indexes)):
# 支持 png jpg 格式的图片.
txtFile = f'{index[:index.rfind(".")]}.txt'
stem = index[:index.rfind(".")]
# 读取图像的宽和高
try:
im = cv2.imread(os.path.join(originImagesDir, index))
height, width, _ = im.shape
except Exception as e:
print(f'{os.path.join(originImagesDir, index)} read error.\nerror:{e}')
# 添加图像的信息
if not os.path.exists(os.path.join(originLabelsDir, txtFile)):
# 如没标签,跳过,只保留图片信息.
continue
dataset['images'].append({'file_name': index,
'id': stem,
'width': width,
'height': height})
with open(os.path.join(originLabelsDir, txtFile), 'r') as fr:
labelList = fr.readlines()
for label in labelList:
label = label.strip().split()
x = float(label[1])
y = float(label[2])
w = float(label[3])
h = float(label[4])
# convert x,y,w,h to x1,y1,x2,y2
H, W, _ = im.shape
x1 = (x - w / 2) * W
y1 = (y - h / 2) * H
x2 = (x + w / 2) * W
y2 = (y + h / 2) * H
# 标签序号从0开始计算, coco2017数据集标号混乱,不管它了。
cls_id = int(label[0])
width = max(0, x2 - x1)
height = max(0, y2 - y1)
dataset['annotations'].append({
'area': width * height,
'bbox': [x1, y1, width, height],
'category_id': cls_id,
'id': ann_id_cnt,
'image_id': stem,
'iscrowd': 0,
# mask, 矩形是从左上角点按顺时针的四个顶点
'segmentation': [[x1, y1, x2, y1, x2, y2, x1, y2]]
})
ann_id_cnt += 1
# 保存结果
with open(arg.save_path, 'w') as f:
json.dump(dataset, f)
print('Save annotation to {}'.format(arg.save_path))
if __name__ == "__main__":
yolo2coco(arg)
2.使用训练好的权重进行验证获得predictions.json
val.py
import warnings
warnings.filterwarnings('ignore')
from ultralytics import RTDETR
# 加载权重
model = RTDETR("/home/root1/data/miya/yolov8/uav/runs/train/rt8/weights/best.pt")
# 数据集yaml文件路径
model.val(data='/home/root1/data/miya/yolov8/data/uavdt/uavdt.yaml',
split='val',
imgsz=1024,
batch=1,
workers=8,
device='0',
# rect=False,
save_json=True, # if you need to cal coco metrice
project='runs/val',
name='exp',
)
运行结果如下,因为我只是测试了一下代码是否正常运行,所以这里使用第一轮的权重,所以可以看到指标都很低
3.使用获得的data.json和predictions.json将YOLO指标转换成COCO指标
get_COCO_metrice.py
import argparse
from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval
from tidecv import TIDE, datasets
def parse_opt():
parser = argparse.ArgumentParser()
# data.json的路径
parser.add_argument('--anno_json', type=str, default='data.json', help='training model path')
# predictions.json的路径
parser.add_argument('--pred_json', type=str, default='runs/val/exp17/predictions.json', help='data yaml path')
return parser.parse_known_args()[0]
if __name__ == '__main__':
opt = parse_opt()
anno_json = opt.anno_json
pred_json = opt.pred_json
anno = COCO(anno_json) # init annotations api
pred = anno.loadRes(pred_json) # init predictions api
eval = COCOeval(anno, pred, 'bbox')
eval.evaluate()
eval.accumulate()
eval.summarize()
tide = TIDE()
tide.evaluate_range(datasets.COCO(anno_json), datasets.COCOResult(pred_json), mode=TIDE.BOX)
tide.summarize()
tide.plot(out_dir='result')
运行一下,指标为0的原因就是因为我用了第一轮的权重,本身验证的就很低