Labelme加载AI（Segment-Anything）模型进行图像标注

news2024/11/16 1:47:39

labelme是使用python写的基于QT的跨平台图像标注工具，可用来标注分类、检测、分割、关键点等常见的视觉任务，支持VOC格式和COCO等的导出，代码简单易读，是非常利用上手的良心工具。
在这里插入图片描述
第一步：
下载源码进行安装。

git clone https://github.com/wkentaro/labelme.git
cd labelme
pip install -e .

第二步：
找到源码所在路径进行修改。
（1）打开labelme/labelme/ai/init.py，源码如下：

MODELS = [
    Model(
        name="Segment-Anything (speed)",
        encoder_weight=Weight(
            url="https://github.com/wkentaro/labelme/releases/download/sam-20230416/sam_vit_b_01ec64.quantized.encoder.onnx",  # NOQA
            md5="80fd8d0ab6c6ae8cb7b3bd5f368a752c",
        ),
        decoder_weight=Weight(
            url="https://github.com/wkentaro/labelme/releases/download/sam-20230416/sam_vit_b_01ec64.quantized.decoder.onnx",  # NOQA
            md5="4253558be238c15fc265a7a876aaec82",
        ),
    ),
    Model(
        name="Segment-Anything (balanced)",
        encoder_weight=Weight(
            url="https://github.com/wkentaro/labelme/releases/download/sam-20230416/sam_vit_l_0b3195.quantized.encoder.onnx",  # NOQA
            md5="080004dc9992724d360a49399d1ee24b",
        ),
        decoder_weight=Weight(
            url="https://github.com/wkentaro/labelme/releases/download/sam-20230416/sam_vit_l_0b3195.quantized.decoder.onnx",  # NOQA
            md5="851b7faac91e8e23940ee1294231d5c7",
        ),
    ),
    Model(
        name="Segment-Anything (accuracy)",
        encoder_weight=Weight(
            url="https://github.com/wkentaro/labelme/releases/download/sam-20230416/sam_vit_h_4b8939.quantized.encoder.onnx",  # NOQA
            md5="958b5710d25b198d765fb6b94798f49e",
        ),
        decoder_weight=Weight(
            url="https://github.com/wkentaro/labelme/releases/download/sam-20230416/sam_vit_h_4b8939.quantized.decoder.onnx",  # NOQA
            md5="a997a408347aa081b17a3ffff9f42a80",
        ),
    ),
]

（2）在labelme/labelme/文件夹下自建一个文件夹model_file。
（3）依次输入以下几个网址下载onnx到model_file文件目录。

https://github.com/wkentaro/labelme/releases/download/sam-20230416/sam_vit_b_01ec64.quantized.encoder.onnx
https://github.com/wkentaro/labelme/releases/download/sam-20230416/sam_vit_b_01ec64.quantized.decoder.onnx

https://github.com/wkentaro/labelme/releases/download/sam-20230416/sam_vit_l_0b3195.quantized.encoder.onnx
https://github.com/wkentaro/labelme/releases/download/sam-20230416/sam_vit_l_0b3195.quantized.decoder.onnx

https://github.com/wkentaro/labelme/releases/download/sam-20230416/sam_vit_h_4b8939.quantized.encoder.onnx
https://github.com/wkentaro/labelme/releases/download/sam-20230416/sam_vit_h_4b8939.quantized.decoder.onnx

在这里插入图片描述
（4）修改labelme/labelme/ai/init.py，代码如下：

import collections

from .models.segment_anything import SegmentAnythingModel  # NOQA


Model = collections.namedtuple(
    "Model", ["name", "encoder_weight", "decoder_weight"]
)

Weight = collections.namedtuple("Weight", ["url", "md5"])

# MODELS = [
#     Model(
#         name="Segment-Anything (speed)",
#         encoder_weight=Weight(
#             url="https://github.com/wkentaro/labelme/releases/download/sam-20230416/sam_vit_b_01ec64.quantized.encoder.onnx",  # NOQA
#             md5="80fd8d0ab6c6ae8cb7b3bd5f368a752c",
#         ),
#         decoder_weight=Weight(
#             url="https://github.com/wkentaro/labelme/releases/download/sam-20230416/sam_vit_b_01ec64.quantized.decoder.onnx",  # NOQA
#             md5="4253558be238c15fc265a7a876aaec82",
#         ),
#     ),
#     Model(
#         name="Segment-Anything (balanced)",
#         encoder_weight=Weight(
#             url="https://github.com/wkentaro/labelme/releases/download/sam-20230416/sam_vit_l_0b3195.quantized.encoder.onnx",  # NOQA
#             md5="080004dc9992724d360a49399d1ee24b",
#         ),
#         decoder_weight=Weight(
#             url="https://github.com/wkentaro/labelme/releases/download/sam-20230416/sam_vit_l_0b3195.quantized.decoder.onnx",  # NOQA
#             md5="851b7faac91e8e23940ee1294231d5c7",
#         ),
#     ),
#     Model(
#         name="Segment-Anything (accuracy)",
#         encoder_weight=Weight(
#             url="https://github.com/wkentaro/labelme/releases/download/sam-20230416/sam_vit_h_4b8939.quantized.encoder.onnx",  # NOQA
#             md5="958b5710d25b198d765fb6b94798f49e",
#         ),
#         decoder_weight=Weight(
#             url="https://github.com/wkentaro/labelme/releases/download/sam-20230416/sam_vit_h_4b8939.quantized.decoder.onnx",  # NOQA
#             md5="a997a408347aa081b17a3ffff9f42a80",
#         ),
#     ),
# ]


MODELS = [
    Model(
        name="Segment-Anything (speed)",
        encoder_weight=Weight(
            url="E:\labelme\labelme\model_file\sam_vit_b_01ec64.quantized.encoder.onnx",  # NOQA
            md5="80fd8d0ab6c6ae8cb7b3bd5f368a752c",
        ),
        decoder_weight=Weight(
            url="E:\labelme\labelme\model_file\sam_vit_b_01ec64.quantized.decoder.onnx",  # NOQA
            md5="4253558be238c15fc265a7a876aaec82",
        ),
    ),
    Model(
        name="Segment-Anything (balanced)",
        encoder_weight=Weight(
            url="E:\labelme\labelme\model_file\sam_vit_l_0b3195.quantized.encoder.onnx",  # NOQA
            md5="080004dc9992724d360a49399d1ee24b",
        ),
        decoder_weight=Weight(
            url="E:\labelme\labelme\model_file\sam_vit_l_0b3195.quantized.decoder.onnx",  # NOQA
            md5="851b7faac91e8e23940ee1294231d5c7",
        ),
    ),
    Model(
        name="Segment-Anything (accuracy)",
        encoder_weight=Weight(
            url="E:\labelme\labelme\model_file\sam_vit_h_4b8939.quantized.encoder.onnx",  # NOQA
            md5="958b5710d25b198d765fb6b94798f49e",
        ),
        decoder_weight=Weight(
            url="E:\labelme\labelme\model_file\sam_vit_h_4b8939.quantized.decoder.onnx",  # NOQA
            md5="a997a408347aa081b17a3ffff9f42a80",
        ),
    ),
]

（5）修改labelme/labelme/widgets/canvas.py，代码如下：

    def initializeAiModel(self, name):
        if name not in [model.name for model in labelme.ai.MODELS]:
            raise ValueError("Unsupported ai model: %s" % name)
        model = [model for model in labelme.ai.MODELS if model.name == name][0]

        if self._ai_model is not None and self._ai_model.name == model.name:
            logger.debug("AI model is already initialized: %r" % model.name)
        else:
            logger.debug("Initializing AI model: %r" % model.name)
            self._ai_model = labelme.ai.SegmentAnythingModel(
                name=model.name,
                # encoder_path=gdown.cached_download(
                #     url=model.encoder_weight.url,
                #     md5=model.encoder_weight.md5,
                # ),
                # decoder_path=gdown.cached_download(
                #     url=model.decoder_weight.url,
                #     md5=model.decoder_weight.md5,
                # ),
                encoder_path=model.encoder_weight.url,
                decoder_path=model.decoder_weight.url,
            )

        self._ai_model.set_image(
            image=labelme.utils.img_qt_to_arr(self.pixmap.toImage())
        )