为什么标注得到的坐标值会超过图像边界

[Ggstuddu]

Search before asking

• I have searched the X-AnyLabeling Docs and issues and found no similar questions.

Question

我使用2.5.3版本的anylabeling进行标注，发觉通过软件转换后的yolo标签有出现坐标值为-1的情况，然后通过排查发觉是在标注的时候得到的json标签文件里的坐标就有问题，得到的坐标（640.4121557454889）超过了图像的边界（640）。
我尝试手动调整一下框的位置，但是更新后的json标签文件仍存在坐标越界问题，请问这个情况应该如何解决？
以下是一个有问题的完整的json标签文件：

{
  "version": "2.5.3",
  "flags": {},
  "shapes": [
    {
      "kie_linking": [],
      "label": "defect_dent",
      "score": null,
      "points": [
        [
          478.68345323741005,
          459.56834532374097
        ],
        [
          512.0863309352518,
          459.56834532374097
        ],
        [
          512.0863309352518,
          484.59712230215825
        ],
        [
          478.68345323741005,
          484.59712230215825
        ]
      ],
      "group_id": null,
      "description": null,
      "difficult": false,
      "shape_type": "rectangle",
      "flags": {},
      "attributes": {}
    },
    {
      "kie_linking": [],
      "label": "defect_dent",
      "score": null,
      "points": [
        [
          63.83950617283952,
          28.33333333333339
        ],
        [
          153.64197530864203,
          28.33333333333339
        ],
        [
          153.64197530864203,
          106.33333333333337
        ],
        [
          63.83950617283952,
          106.33333333333337
        ]
      ],
      "group_id": null,
      "description": null,
      "difficult": false,
      "shape_type": "rectangle",
      "flags": {},
      "attributes": {}
    },
    {
      "kie_linking": [],
      "label": "defect_dent",
      "score": null,
      "points": [
        [
          362.90123456790127,
          4.0
        ],
        [
          458.0,
          4.0
        ],
        [
          458.0,
          49.320987654321044
        ],
        [
          362.90123456790127,
          49.320987654321044
        ]
      ],
      "group_id": null,
      "description": null,
      "difficult": false,
      "shape_type": "rectangle",
      "flags": {},
      "attributes": {}
    },
    {
      "kie_linking": [],
      "label": "defect_dent",
      "score": null,
      "points": [
        [
          557.3170731707318,
          372.19512195121956
        ],
        [
          574.3902439024391,
          372.19512195121956
        ],
        [
          574.3902439024391,
          390.2439024390244
        ],
        [
          557.3170731707318,
          390.2439024390244
        ]
      ],
      "group_id": null,
      "description": "",
      "difficult": false,
      "shape_type": "rectangle",
      "flags": {},
      "attributes": {}
    },
    {
      "kie_linking": [],
      "label": "defect_dirt",
      "score": null,
      "points": [
        [
          372.3570750237418,
          529.7454890788223
        ],
        [
          640.4311490978159,
          529.7454890788223
        ],
        [
          640.4311490978159,
          640.4121557454889
        ],
        [
          372.3570750237418,
          640.4121557454889
        ]
      ],
      "group_id": null,
      "description": "",
      "difficult": false,
      "shape_type": "rectangle",
      "flags": {},
      "attributes": {}
    },
    {
      "kie_linking": [],
      "label": "defect_dent",
      "score": null,
      "points": [
        [
          462.28395061728395,
          556.1111111111111
        ],
        [
          491.2962962962963,
          556.1111111111111
        ],
        [
          491.2962962962963,
          580.8024691358025
        ],
        [
          462.28395061728395,
          580.8024691358025
        ]
      ],
      "group_id": null,
      "description": "",
      "difficult": false,
      "shape_type": "rectangle",
      "flags": {},
      "attributes": {}
    },
    {
      "kie_linking": [],
      "label": "defect_long_strip",
      "score": null,
      "points": [
        [
          146.23456790123458,
          23.395061728395117
        ],
        [
          580.1851851851852,
          23.395061728395117
        ],
        [
          580.1851851851852,
          132.6543209876544
        ],
        [
          146.23456790123458,
          132.6543209876544
        ]
      ],
      "group_id": null,
      "description": "",
      "difficult": false,
      "shape_type": "rectangle",
      "flags": {},
      "attributes": {}
    }
  ],
  "imagePath": "new_21.jpg",
  "imageData": null,
  "imageHeight": 640,
  "imageWidth": 640,
  "description": ""
}

Additional

No response

[CVHub520]

Hi, @Ggstuddu:

Thanks for your feedback! Before I can provide a detailed solution, could you tell me are you using the default built-in model provided by X-AnyLabeling, or a custom-trained one?

If you could provide these details, I can better diagnose the root cause and suggest appropriate fixes for the coordinate prediction issue.

[Ggstuddu]

你好，十分感谢你的回复。
我使用的是你们提供的模型，下载链接是：https://github.com/CVHub520/X-AnyLabeling/releases/download/v0.1.0/yolov8s.onnx
在模型自动跑然后运行了之后我也试过手动调整bbox，但更新后的json文件的坐标还是越界。

[CVHub520]

The official built-in model is trained on the COCO dataset weights. As you provided in the label file example above, it includes categories such as defect_dent and defect_dirt. To my knowledge, these are not part of the 80 categories defined by COCO. Could there be some misunderstanding here?

[Ggstuddu]

对不起，我记混了，记成另外一个项目了，这个越界的框应该是我自己手动框的，或者是通过yolo模型预测得到的txt导入得到的，具体是哪一个我不太记得，因为如果yolo模型预测的不准的话我会自己手动框一下。
并没有用到你们提供的onnx文件，很抱歉我记错了。

[CVHub520]

If you manually draw a rectangle box, the maximum coordinate value is (image_height-1, image_width-1). Specifically, you can load any new image and try drawing a rectangle box.

Therefore, this is most likely due to a deviation in the coordinates output by your custom model. But, don't worry, you can just add a simple truncation logic here.

[Ggstuddu]

感谢您的建议，我觉得应该是可行的，我明天会添加截断功能并且测试一下。

[Ggstuddu]

今天再进行了测试，发觉问题并没有完全解决：
为了确保标签坐标不会越界，首先要确保标签数据导入并转换成json文件的时候不会越界，目前基于已有的yolo模型获得json标签文件有两种方式：
①通过anylabeling提供的自动标注功能：在这个情况为了确保不要越界，需要在anylabeling/services/auto_labeling/base/yolo.py的create_rectangle_shape()函数添加截断功能；
②通过模型预测得到标签txt文件，然后通过“导入yolo水平框”功能导入：在这个情况为了确保不要越界，需要在anylabeling/views/labeling/label_converter.py的yolo_to_custom()函数添加截断功能。
经测试，自己之前确实会在导入和转换成json文件时直接产生越界的bbox坐标，但在对应的区域添加了截断功能后导入生成的json文件的坐标已经不会有越界的问题了。
但是后面发现在将bbox往图像边界拖拽时仍有可能会在json标签文件那里产生越界的坐标值，可能也需要在拖拽bbox的代码添加截断的功能。

[CVHub520]

Regarding the possibility of generating out-of-bounds coordinate values in the JSON label file when dragging a rectangle towards the canvas boundary, would you be able to record a video demonstrating this? I have tested this multiple times and have not encountered the situation.

[Ggstuddu]

我在家里的电脑复现了一下，家里用的anylabeling的版本是2.4.4，仍有出现拖拽后越界的情况，以下是视频：

2025-02-24-20-53-59.mp4

[CVHub520]

Thank you for the detailed feedback. It does appear there is an overflow of coordinate values.