我的需求是只保留人（和没有人的副样本），抽样，其他忽略

import json
import os
import shutil
import random

def coco_to_yolo_bbox(image_width, image_height, bbox):
    """
    Convert bbox from COCO format to YOLO format.
    COCO format: [top_left_x, top_left_y, width, height]
    YOLO format: [x_center, y_center, width, height] (normalized)
    """
    x_tl, y_tl, width, height = bbox
    x_center = (x_tl + width / 2.0) / image_width
    y_center = (y_tl + height / 2.0) / image_height
    width /= image_width
    height /= image_height
    return x_center, y_center, width, height

def filter_and_convert_annotations(input_annotation_file, output_annotation_dir, output_image_dir, src_image_dir, num_positive_samples=None, num_negative_samples=None):
    os.makedirs(output_image_dir, exist_ok=True)
    os.makedirs(output_annotation_dir, exist_ok=True)

    with open(input_annotation_file, 'r') as f:
        coco = json.load(f)

    # 获取类别为 "person" 的 ID
    person_category_id = [cat['id'] for cat in coco['categories'] if cat['name'] == 'person'][0]

    # 过滤注释，只保留人类并且有边界框信息的注释
    person_annotations = [anno for anno in coco['annotations'] if anno['category_id'] == person_category_id and 'bbox' in anno]

    # 获取所有含有"人"的图像 ID
    person_image_ids = set(anno['image_id'] for anno in person_annotations)

    # 过滤图片，只保留含有人的图片
    person_images = [img for img in coco['images'] if img['id'] in person_image_ids]
    
    # 设置正样本的数量
    if num_positive_samples is None:
        num_positive_samples = len(person_images)
    else:
        num_positive_samples = min(num_positive_samples, len(person_images))

    # 随机抽取指定数量的正样本
    person_images = random.sample(person_images, num_positive_samples)
    positive_image_ids = set(img['id'] for img in person_images)  # 获取选中的正样本图片 ID

    # 过滤注释，只保留选中的正样本图片的注释
    person_annotations = [anno for anno in person_annotations if anno['image_id'] in positive_image_ids]

    # 获取没有人的图片 ID
    all_image_ids = set(img['id'] for img in coco['images'])
    nonperson_image_ids = list(all_image_ids - person_image_ids)  # 转换为列表

    # 设置负样本的数量
    if num_negative_samples is None:
        num_negative_samples = len(nonperson_image_ids)
    else:
        num_negative_samples = min(num_negative_samples, len(nonperson_image_ids))

    # 随机抽取指定数量的负样本
    negative_image_ids = random.sample(nonperson_image_ids, num_negative_samples)

    # 过滤图片，只保留负样本
    negative_images = [img for img in coco['images'] if img['id'] in negative_image_ids]

    # 将图像和标注转换为YOLO格式
    for img in person_images:
        img_id = img['id']
        img_width = img['width']
        img_height = img['height']
        img_file_name = img['file_name']
        img_annotations = [anno for anno in person_annotations if anno['image_id'] == img_id]

        yolo_annotations = []
        for anno in img_annotations:
            bbox = anno['bbox']
            x_center, y_center, width, height = coco_to_yolo_bbox(img_width, img_height, bbox)
            yolo_annotation = f'0 {x_center} {y_center} {width} {height}'
            # yolo_annotation = f'{person_category_id} {x_center} {y_center} {width} {height}'
            yolo_annotations.append(yolo_annotation)

        annotation_file_name = os.path.splitext(img_file_name)[0] + '.txt'
        with open(os.path.join(output_annotation_dir, annotation_file_name), 'w') as f:
            f.write('\n'.join(yolo_annotations))

    # 为负样本创建空标注文件
    for img in negative_images:
        img_file_name = img['file_name']
        annotation_file_name = os.path.splitext(img_file_name)[0] + '.txt'
        open(os.path.join(output_annotation_dir, annotation_file_name), 'w').close()

    # 复制图像
    selected_image_filenames = [img['file_name'] for img in person_images + negative_images]

    total_files = len(selected_image_filenames)
    for i, filename in enumerate(selected_image_filenames):
        src_path = os.path.join(src_image_dir, filename)
        dst_path = os.path.join(output_image_dir, filename)
        if not os.path.exists(dst_path):  # 确保不重复拷贝
            shutil.copyfile(src_path, dst_path)
        if i % 100 == 0:  # 每拷贝100张图片打印一次进度
            print(f"Copied {i}/{total_files} images")

# 设置输入输出路径
filter_and_convert_annotations(
    '/path/coco_2017/annotations_trainval/instances_train2017.json',
    '/path/coco_2017_yolo/train_label',
    '/path/coco_2017_yolo/train',
    '/path/coco_2017/train2017',
    num_positive_samples=4000,
    num_negative_samples=400
)

filter_and_convert_annotations(
    '/path/coco_2017/annotations_trainval/instances_val2017.json',
    '/path/coco_2017_yolo/val_label',
    '/path/coco_2017_yolo/val',
    '/path/coco_2017/val2017',
    num_positive_samples=400,
    num_negative_samples=40
)