摘要

V3Det：一个庞大的词汇视觉检测数据集，在大量真实世界图像上具有精确注释的边界框，其包含13029个类别中的245k个图像（比LVIS大10倍），数据集已经开源！

图片的数量比COCO多一些，类别种类比较多！数据集大小由33G，数据集标注格式和COCO一致！
论文链接：https://arxiv.org/abs/2304.03752

这个数据集最大的特点就是类别多，还有些千奇百怪不可描述的图片！

下载V3Det的标注文件

官方提供了两种下载方式，见：https://v3det.openxlab.org.cn/download
第一种，点击左侧的链接，将其中的文件都下载下来！

v3det_2023_v1_train.json和v3det_2023_v1_val.json是数据集！
v3det_image_download.py是下载图片的脚本。
category_name_13204_v3det_2023_v1.txt 是类别！
第二种下载方式如下：

采用命令行，注册后输入密钥就能下载！下载下来的文件和第一种下载方式的文件一样，都没有图像，只能运行脚本下载图片！

下载图片的脚本

由于总所周知的原因不太好链接，多试几次，总有成功的时候。

import io
import argparse
import concurrent.futures
import json
import os
import time
import urllib.error
import urllib.request

from tqdm import tqdm

parser = argparse.ArgumentParser()
parser.add_argument("--output_folder", type=str, default="V3Det")
parser.add_argument("--max_retries", type=int, default=3)
parser.add_argument("--max_workers", type=int, default=16)
args = parser.parse_args()
headers = {'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/113.0.0.0 Safari/537.36'}


def cache(response):
    f = io.BytesIO()
    block_sz = 8192
    while True:
        buffer = response.read(block_sz)
        if not buffer:
            break
        f.write(buffer)
    return f

def download_image(url, path, timeout):
    result = {
        "status": "",
        "url": url,
        "path": path,
    }
    cnt = 0
    while True:
        try:
            response = urllib.request.urlopen(urllib.request.Request(url=url, headers=headers), timeout=timeout)
            image_path = os.path.join(args.output_folder, path)
            os.makedirs(os.path.dirname(image_path), exist_ok=True)
            f = cache(response)
            with open(image_path, "wb") as fp:
                fp.write(f.getvalue())
            result["status"] = "success"
        except Exception as e:
            if not isinstance(e, urllib.error.HTTPError):
                cnt += 1
                if cnt <= args.max_retries:
                    continue
            if isinstance(e, urllib.error.HTTPError):
                result["status"] = "expired"
            else:
                result["status"] = "timeout"
        break
    return result


def main():
    start = time.time()
    if os.path.exists(args.output_folder) and os.listdir(args.output_folder):
        try:
            c = input(
                f"'{args.output_folder}' already exists and is not an empty directory, continue? (y/n) "
            )
            if c.lower() not in ["y", "yes"]:
                exit(0)
        except KeyboardInterrupt:
            exit(0)
    if not os.path.exists(args.output_folder):
        os.makedirs(args.output_folder)
    image_folder_path = os.path.join(args.output_folder, "images")
    record_path = os.path.join(args.output_folder, "records.json")
    record = {'success': [], 'expired': [], 'timeout': []}
    if os.path.isfile(record_path):
        try:
            with open(record_path, encoding="utf8") as f:
                record['success'] = json.load(f)['success']
        except:
            pass
    if not os.path.exists(image_folder_path):
        os.makedirs(image_folder_path)

    list_url = 'https://raw.githubusercontent.com/V3Det/v3det_resource/main/resource/download_list.txt'
    response = urllib.request.urlopen(urllib.request.Request(url=list_url, headers=headers), timeout=100)
    url_list = [url for url in response.read().decode('utf-8').split('\n') if len(url) > 0]
    image2url = {}
    for url in url_list:
        response = urllib.request.urlopen(urllib.request.Request(url=url, headers=headers), timeout=100)
        image2url.update(eval(response.read().decode('utf-8')))

    data = []
    rec_suc = set(record['success'])
    for image, url in image2url.items():
        if image not in rec_suc:
            data.append((url, image))
    with tqdm(total=len(data)) as pbar:
        with concurrent.futures.ThreadPoolExecutor(max_workers=args.max_workers) as executor:
            # Submit up to `chunk_size` tasks at a time to avoid too many pending tasks.
            chunk_size = min(5000, args.max_workers * 500)
            for i in range(0, len(data), chunk_size):
                futures = [
                    executor.submit(download_image, url, path, 10)
                    for url, path in data[i: i + chunk_size]
                ]
                for future in concurrent.futures.as_completed(futures):
                    r = future.result()
                    record[r["status"]].append(r["path"])
                    pbar.update(1)
                with open(record_path, "w", encoding="utf8") as f:
                    json.dump(record, f, indent=2)

    end = time.time()
    print(f"consuming time {end - start:.1f} sec")
    print(f"{len(record['success'])} images downloaded.")
    print(f"{len(record['timeout'])} urls failed due to request timeout.")
    print(f"{len(record['expired'])} urls failed due to url expiration.")
    if len(record['success']) == len(image2url):
        os.remove(record_path)
        print('All images have been downloaded!')
    else:
        print('Please run this file again to download failed image!')


if __name__ == "__main__":
    main()

V3Det转Yolo

V3Det的标注文件和COCO是一致的！

import json
import os
import shutil
from pathlib import Path
import numpy as np
from tqdm import tqdm


def make_folders(path='../out/'):
    # Create folders

    if os.path.exists(path):
        shutil.rmtree(path)  # delete output folder
    os.makedirs(path)  # make new output folder
    os.makedirs(path + os.sep + 'labels')  # make new labels folder
    os.makedirs(path + os.sep + 'images')  # make new labels folder
    return path


def convert_coco_json(json_dir='./image_1024/V3Det___V3Det/raw/v3det_2023_v1_val.json',out_dir=None):
    # fn_images = 'out/images/%s/' % Path(json_file).stem.replace('instances_', '')  # folder name
    os.makedirs(out_dir,exist_ok=True)
    # os.makedirs(fn_images,exist_ok=True)
    with open(json_dir) as f:
            data = json.load(f)
    print(out_dir)
    # Create image dict
    images = {'%g' % x['id']: x for x in data['images']}

        # Write labels file
    for x in tqdm(data['annotations'], desc='Annotations %s' % json_dir):
        if x['iscrowd']:
            continue

        img = images['%g' % x['image_id']]
        h, w, f = img['height'], img['width'], img['file_name']
        file_path='coco/'+out_dir.split('/')[-2]+"/"+f
        # The Labelbox bounding box format is [top left x, top left y, width, height]
        box = np.array(x['bbox'], dtype=np.float64)
        box[:2] += box[2:] / 2  # xy top-left corner to center
        box[[0, 2]] /= w  # normalize x
        box[[1, 3]] /= h  # normalize y

        if (box[2] > 0.) and (box[3] > 0.):  # if w > 0 and h > 0
            with open(out_dir + Path(f).stem + '.txt', 'a') as file:
                file.write('%g %.6f %.6f %.6f %.6f\n' % (x['category_id'] - 1, *box))




convert_coco_json(json_dir='./image_1024/V3Det___V3Det/raw/v3det_2023_v1_val.json',out_dir='out/labels/val/')
convert_coco_json(json_dir='./image_1024/V3Det___V3Det/raw/v3det_2023_v1_train.json',out_dir='out/labels/train/')

复制图片到指定目录

将图片放到和Label同级的images文件夹

import glob
import os
import shutil

image_paths = glob.glob('V3Det/images/*/*.jpg')

dir_imagepath = {}

for image_path in image_paths:
    image_key = image_path.replace('\\', '/').split('/')[-1].split('.')[0]
    dir_imagepath[image_key] = image_path

os.makedirs('out/images/train',exist_ok=True)
os.makedirs('out/images/val',exist_ok=True)


def txt_2_image(txt_dir='out/labels/train/', out_path='out/images/train'):
    txt_paths = glob.glob(txt_dir + '*.txt')
    for txt in txt_paths:
        txt_key = txt.replace('\\', '/').split('/')[-1].split('.')[0]
        if txt_key in dir_imagepath:
            image_path = dir_imagepath[txt_key]
            shutil.copy(image_path, out_path)
        else:
            os.remove(txt)


txt_2_image(txt_dir='out/labels/train/', out_path='out/images/train')
txt_2_image(txt_dir='out/labels/val/', out_path='out/images/val')

生成类别

找到类别文件，生成YoloV5或V8的类别格式，如下图：

代码如下：

with open('image_1024/V3Det___V3Det/raw/category_name_13204_v3det_2023_v1.txt','r') as files:
    list_class=files.readlines()
    for i, c in enumerate(list_class):
        print(str(i)+": "+c.replace('\n',''))

将生成的类别复制到YoloV8或者V5的数据集配置文件中！

总结

这个数据集比COCO数据集大一些，种类更加丰富，可以使用这个数据集训练，做预训练权重！

经测验，使用V3Det训练的模型做预训练权重，训练COCO可以提升1MAp！