{
  "id": 304293,
  "title": "yolov5 map vs pycocotools map difference",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/304293",
  "author_name": "jaidmin",
  "post_date": "2022-01-31T14:59:18.377000",
  "votes": 4,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hey everyone, </p>\n<p>I've been noticing some drastic mAP 0.5:0.95 differences between the yolov5 evaluation and pycocotools. This seems to be a <a href=\"https://github.com/ultralytics/yolov5/issues/2258\" target=\"_blank\">known issue</a> but in the issue discussion they talk about small difference like 1% while I see 5-8 % on some folds.</p>\n<p>I will double check my code but am reasonably sure that its correct, I use <code>python val.py --save-txt --save-conf</code> with yolov5, and then parse the generated txt files, change the box format to coco and evaluate with pycocotools.</p>\n<p>Has anyone seen something like this before? I assume its better to trust the pycocotools results</p>",
  "messages": [
    {
      "id": 1670480,
      "postDate": "2022-01-31T14:59:18.377Z",
      "content": "<p>Hey everyone, </p>\n<p>I've been noticing some drastic mAP 0.5:0.95 differences between the yolov5 evaluation and pycocotools. This seems to be a <a href=\"https://github.com/ultralytics/yolov5/issues/2258\" target=\"_blank\">known issue</a> but in the issue discussion they talk about small difference like 1% while I see 5-8 % on some folds.</p>\n<p>I will double check my code but am reasonably sure that its correct, I use <code>python val.py --save-txt --save-conf</code> with yolov5, and then parse the generated txt files, change the box format to coco and evaluate with pycocotools.</p>\n<p>Has anyone seen something like this before? I assume its better to trust the pycocotools results</p>",
      "rawMarkdown": "Hey everyone, \n\nI've been noticing some drastic mAP 0.5:0.95 differences between the yolov5 evaluation and pycocotools. This seems to be a [known issue](https://github.com/ultralytics/yolov5/issues/2258) but in the issue discussion they talk about small difference like 1% while I see 5-8 % on some folds.\n\nI will double check my code but am reasonably sure that its correct, I use `python val.py --save-txt --save-conf` with yolov5, and then parse the generated txt files, change the box format to coco and evaluate with pycocotools.\n\nHas anyone seen something like this before? I assume its better to trust the pycocotools results",
      "votes": 4
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1670480": "Hey everyone, \n\nI've been noticing some drastic mAP 0.5:0.95 differences between the yolov5 evaluation and pycocotools. This seems to be a [known issue](https://github.com/ultralytics/yolov5/issues/2258) but in the issue discussion they talk about small difference like 1% while I see 5-8 % on some folds.\n\nI will double check my code but am reasonably sure that its correct, I use `python val.py --save-txt --save-conf` with yolov5, and then parse the generated txt files, change the box format to coco and evaluate with pycocotools.\n\nHas anyone seen something like this before? I assume its better to trust the pycocotools results"
  }
}