YOLO11+SAHI输出指标

首先使用sahi结合你的模型进行检测，这里面wh还有重叠率决定了切片数量和检测用时。

先是定义模型，将权重传入，并设置置信度

detection_model = AutoDetectionModel.from_pretrained(
    model_type="ultralytics",
    model_path=model_path,
    confidence_threshold=0.3,
    device='cuda:0',  # or 'cuda:0'
)

使用get_sliced_prediction函数得到推理结果

 from sahi.predict import  get_sliced_prediction  
 result = get_sliced_prediction(
            imagepath,
            detection_model,
            slice_height=640,
            slice_width=640,
            overlap_height_ratio=0.2,
            overlap_width_ratio=0.2,
        )

将结果转化为coco所需的json格式，这里需要与验证集的一致，我理解是imge_id

一致，可以通过下面函数传给他image_id,比如验证集json使用的是文件名，那么这里就要传给他文件名

pre = result.to_coco_predictions(image_id=imagepath)

1.之后是生成json的完整代码

import warnings
warnings.filterwarnings('ignore')
from ultralytics import YOLO
from sahi import AutoDetectionModel
from sahi.predict import predict, get_sliced_prediction
import os
import json
from tqdm import tqdm
#模型路径
model_path = 'runs/train/exp9/weights/best.pt'
#验证集路径
image_dir = 'ultralytics/assets'
def getjson(imagepath, detection_model):
    try:
        predictions = []
        result = get_sliced_prediction(
            imagepath,
            detection_model,
            slice_height=640,
            slice_width=640,
            overlap_height_ratio=0.2,
            overlap_width_ratio=0.2,
        )
        id = os.path.splitext(os.path.basename(imagepath))[0]
        pre = result.to_coco_predictions(image_id=id)
        for prediction in pre:
            new_prediction = {
                "image_id": prediction["image_id"],
                "category_id": prediction["category_id"],
                "bbox": prediction["bbox"],
                "score": prediction["score"]
            }
            predictions.append(new_prediction)
        return predictions
    except Exception as e:
        print(f"处理图片 {imagepath} 时出错: {e}")
        return []



detection_model = AutoDetectionModel.from_pretrained(
    model_type="ultralytics",
    model_path=model_path,
    confidence_threshold=0.3,
    device='cuda:0',  # or 'cuda:0'
)


all_predictions = []

# 获取所有 .jpg 图片路径
jpg_images = [os.path.join(image_dir, img) for img in os.listdir(image_dir) if img.endswith('.jpg')]

# 使用 tqdm 包装循环，显示进度条
for image_path in tqdm(jpg_images, desc="处理图片进度"):
    result = getjson(image_path, detection_model)
    all_predictions += result



with open('results5.json', 'w', encoding='utf-8') as f:
    json.dump(all_predictions, f, ensure_ascii=False, indent=4)
    

2.然后是对测试集生成

import os
import cv2
import json
from tqdm import tqdm
from sklearn.model_selection import train_test_split
import argparse

# 输入测试集类别
classes = ["mini", "serious"]
parser = argparse.ArgumentParser()
# 在default或者--给出测试集图片和标签路径
parser.add_argument('--image_path', default='/root/autodl-tmp/traffiicaccident/accident/images2',type=str, help="path of images")
parser.add_argument('--label_path', default='/root/autodl-tmp/traffiicaccident/accident/labels2',type=str, help="path of labels .txt")
parser.add_argument('--save_path', default='data.json', type=str, 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)

3.使用上述两段代码生成的json来计算coco指标

import warnings
warnings.filterwarnings('ignore')
import argparse
from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval
from tidecv import TIDE, datasets
#在default或者--给出上面的anno路径和pre路径，分别是测试集的json和预测的json
def parse_opt():
    parser = argparse.ArgumentParser()
    parser.add_argument('--anno_json', type=str, default='/root/autodl-tmp/UAV_smoke_fire/YOLO_UAV/data.json', help='label coco json path')
    parser.add_argument('--pred_json', type=str, default='/root/autodl-tmp/drone/ultralytics-yolo11-main/runs/val/exp8/predictions.json', help='pred coco json path')
    #autodl-tmp/drone/ultralytics-yolo11-main/runs/val/exp8/predictions.json
    #sahi，results.json
    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')

最后得到结果