利用SAM实现自动标注_sam标注

利用SAM实现自动标注

---

一、下载安装

1.1 下载SAM-Tool和SAM

git clone https://github.com/zhouayi/SAM-Tool.git
git clone https://github.com/facebookresearch/segment-anything.git

1.2 下载SAM模型文件

wget https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth

1.3 安装SAM

进入SAM文件夹：cd segment-anything
安装SAM：pip install -e .

二、配置项目

（1）将SAM工具中的helper文件夹移动到SAM主文件夹中：cp -r helpers/ …/segment-anything/
（2）构建数据文件夹格式为dataset_path/images和dataset_path/embeddings(images里用来放图片)

三、提取信息

（1）提取图片embedding：python helpers/extract_embeddings.py --checkpoint-path sam_vit_h_4b8939.pth --dataset-folder 你的数据集地址 --device cuda
（2）运行完在embeddings下会有对应的npy文件

四、获得SAM onnx文件

（1）首先环境中要有onnx和onnxruntime(-gpu)
（2）检测图像的文件(可调整后面的图片高/宽)：python helpers/generate_onnx.py --checkpoint-path sam_vit_h_4b8939.pth --onnx-model-path ./sam_onnx.onnx --orig-im-size 720 1280
（3）运行完会有对应的sam_onnx.onnx文件，将其移到SAM工具主文件夹中：cp sam_onnx.onnx …/SAM-Tool/

五、标注

（1）opencv库要删掉opencv-python，改用headless版的：pip install opencv-python-headless
（2）调用标注程序命令为(后面可改标注类别，用逗号隔开)：python segment_anything_annotator.py --onnx-model-path sam_onnx.onnx --dataset-path /home/dxfcv/workspace/sunsirui/label/segment-anything/dataset --categories cement,metal,plastics
（3）标注框出现后开始标注，首先点击目标，然后添加对象，最后选定类别，接下来下一张，标完所有就保存，最后会在数据文件夹下出现annotations.json的coco格式文件
（4）查看标注文件：python cocoviewer.py -i …/segment-anything/dataset/ -a …/segment-anything/dataset/annotations.json

六、格式转换

6.1 json转voc

import json
import xml.etree.ElementTree as ET
import os

jsonPath = "/home/dxfcv/workspace/sunsirui/label/dataset/annotations.json"#改这个输入json
vocPath = "/home/dxfcv/workspace/sunsirui/label/dataset/voc"//# 改这个输出文件夹
with open(jsonPath, 'r') as f:
    data = json.load(f)

info = data["info"]
images = data["images"]
annotations = data["annotations"]
categories = data["categories"]

# 对每个图像处理
for img_data in images:
    # 创建 VOC XML 文件
    xml_file = ET.Element('annotation')
    ET.SubElement(xml_file, 'folder').text = 'VOC'
    ET.SubElement(xml_file, 'filename').text = os.path.basename(img_data["file_name"])
    source = ET.SubElement(xml_file, 'source')
    ET.SubElement(source, 'database').text = 'My Database'
    ET.SubElement(source, 'annotation').text = 'COCO'
    ET.SubElement(source, 'image').text = 'flickr'
    size = ET.SubElement(xml_file, 'size')
    ET.SubElement(size, 'width').text = str(img_data['width'])
    ET.SubElement(size, 'height').text = str(img_data['height'])
    ET.SubElement(size, 'depth').text = '3'
    ET.SubElement(xml_file, 'segmented').text = '0'

    # 查找该图像的所有标注框
    bbox_list = []
    category_ids = []
    for ann_data in annotations:
        if ann_data['image_id'] == img_data['id']:
            bbox = ann_data['bbox']
            bbox_list.append(bbox)
            category_ids.append(ann_data['category_id'])

    # 对每个标注框处理
    for i in range(len(bbox_list)):
        bbox = bbox_list[i]
        category_id = category_ids[i]
        # 转换 COCO 格式到 VOC 格式
        x_min = bbox[0]
        y_min = bbox[1]
        x_max = bbox[0] + bbox[2]
        y_max = bbox[1] + bbox[3]
        class_name = categories[category_id]['name']
        # 创建 VOC XML 标注
        obj = ET.SubElement(xml_file, 'object')
        ET.SubElement(obj, 'name').text = class_name
        ET.SubElement(obj, 'pose').text = 'Unspecified'
        ET.SubElement(obj, 'truncated').text = '0'
        ET.SubElement(obj, 'difficult').text = '0'
        bndbox = ET.SubElement(obj, 'bndbox')
        ET.SubElement(bndbox, 'xmin').text = str(int(x_min))
        ET.SubElement(bndbox, 'ymin').text = str(int(y_min))
        ET.SubElement(bndbox, 'xmax').text = str(int(x_max))
        ET.SubElement(bndbox, 'ymax').text = str(int(y_max))

    # 将 XML 文件保存到 VOC 目标文件夹中
    xml_str = ET.tostring(xml_file)
    with open(os.path.join(vocPath, os.path.basename(img_data["file_name"]).replace('.jpg', '.xml')), 'wb') as f:
        f.write(xml_str)

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66

6.2 json转yolo

import os
import json
from tqdm import tqdm
import argparse
 
 
def convert(size, box):
    '''
    size: 图片的宽和高(w,h)
    box格式: x,y,w,h
    返回值：x_center/image_width y_center/image_height width/image_width height/image_height
    '''
 
    dw = 1. / (size[0])
    dh = 1. / (size[1])
    x = box[0] + box[2] / 2.0
    y = box[1] + box[3] / 2.0
    w = box[2]
    h = box[3]
 
    x = x * dw
    w = w * dw
    y = y * dh
    h = h * dh
    return (x, y, w, h)
 
 
if __name__ == '__main__':
 
    parser = argparse.ArgumentParser()
    parser.add_argument('--json_file', default='/home/dxfcv/workspace/sunsirui/label/dataset/annotations.json',
                        type=str, help="coco file path")
    parser.add_argument('--save_dir', default='/home/dxfcv/workspace/sunsirui/label/dataset/labels', type=str,
                        help="where to save .txt labels")
    arg = parser.parse_args()
 
    data = json.load(open(arg.json_file, 'r'))
 
    # 如果存放txt文件夹不存在，则创建
    if not os.path.exists(arg.save_dir):
        os.makedirs(arg.save_dir)
 
    id_map = {}
 
    # 解析目标类别，也就是 categories 字段，并将类别写入文件 classes.txt 中
    with open(os.path.join(arg.save_dir, 'classes.txt'), 'w') as f:
        for i, category in enumerate(data['categories']):
            f.write(f"{category['name']}\n")
            id_map[category['id']] = i
 
    for img in tqdm(data['images']):
 
        # 解析 images 字段，分别取出图片文件名、图片的宽和高、图片id
        # filename = os.path.join(arg.image_file , img["file_name"])
        filename =  img["file_name"].split('/')[1]
        img_width = img["width"]
        img_height = img["height"]
        img_id = img["id"]
        head, tail = os.path.splitext(filename)
 
        # txt文件名，与对应图片名只有后缀名不一样
        txt_name = head + ".txt"
        f_txt = open(os.path.join(arg.save_dir, txt_name), 'w')
 
        for ann in data['annotations']:
            if ann['image_id'] == img_id:
                box = convert((img_width, img_height), ann["bbox"])
 
                # 写入txt，共5个字段
                f_txt.write("%s %s %s %s %s\n" % (
                    id_map[ann["category_id"]], box[0], box[1], box[2], box[3]))
 
        f_txt.close()

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75