{
  "id": 123985,
  "title": "What's reason of high local mAP and low LB",
  "url": "/competitions/pku-autonomous-driving/discussion/123985",
  "author_name": "",
  "post_date": "2020-01-01T06:01:03.986639300Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I've implemented original centernet on this dataset and achieved 0.114 local mAP but when I submit the predictions I only get LB around 0.02. Playing with different thresholds can affect local mAP but nearly all LB scores are 0.02. I thought once the correct mAP is computed the LB score should be fine. This is the first time I enter a detection competition and couldn't figure out why. Any suggestions would be appreciated. Thanks!</p>\n\n<p>PS: The mAP calculation code is borrowed from public kernels.</p>",
  "messages": [
    {
      "id": "707597",
      "postDate": "01/01/2020 06:01:03",
      "content": "<p>I've implemented original centernet on this dataset and achieved 0.114 local mAP but when I submit the predictions I only get LB around 0.02. Playing with different thresholds can affect local mAP but nearly all LB scores are 0.02. I thought once the correct mAP is computed the LB score should be fine. This is the first time I enter a detection competition and couldn't figure out why. Any suggestions would be appreciated. Thanks!</p>\n\n<p>PS: The mAP calculation code is borrowed from public kernels.</p>",
      "rawMarkdown": "I've implemented original centernet on this dataset and achieved 0.114 local mAP but when I submit the predictions I only get LB around 0.02. Playing with different thresholds can affect local mAP but nearly all LB scores are 0.02. I thought once the correct mAP is computed the LB score should be fine. This is the first time I enter a detection competition and couldn't figure out why. Any suggestions would be appreciated. Thanks!\n\nPS: The mAP calculation code is borrowed from public kernels.",
      "votes": null
    },
    {
      "id": "707921",
      "postDate": "01/01/2020 17:27:19",
      "content": "<p>Sounds like you are overfitting to the training data. What size split are you using for local validation?</p>",
      "rawMarkdown": "Sounds like you are overfitting to the training data. What size split are you using for local validation?",
      "votes": null
    },
    {
      "id": "708298",
      "postDate": "01/02/2020 06:44:26",
      "content": "<p>Thanks for your reply. The problem is solved. I'm not overfitting the training data but using the regression <code>x, y, z</code> values as submitting results while using the coordinates from the output heatmap to calculate local mAP. However, the 3d regression <code>x, y, z</code> doesn't always match the 2d coordinates and this leads to large gap between local mAP and LB score.</p>",
      "rawMarkdown": "Thanks for your reply. The problem is solved. I'm not overfitting the training data but using the regression `x, y, z` values as submitting results while using the coordinates from the output heatmap to calculate local mAP. However, the 3d regression `x, y, z` doesn't always match the 2d coordinates and this leads to large gap between local mAP and LB score.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 707921,
      "author_name": "jackvial",
      "author_url": "",
      "post_date": "01/01/2020 17:27:19",
      "content": "<p>Sounds like you are overfitting to the training data. What size split are you using for local validation?</p>",
      "votes": null,
      "replies": [
        {
          "id": 708298,
          "author_name": "stannnn",
          "author_url": "",
          "post_date": "01/02/2020 06:44:26",
          "content": "<p>Thanks for your reply. The problem is solved. I'm not overfitting the training data but using the regression <code>x, y, z</code> values as submitting results while using the coordinates from the output heatmap to calculate local mAP. However, the 3d regression <code>x, y, z</code> doesn't always match the 2d coordinates and this leads to large gap between local mAP and LB score.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "707597": "I've implemented original centernet on this dataset and achieved 0.114 local mAP but when I submit the predictions I only get LB around 0.02. Playing with different thresholds can affect local mAP but nearly all LB scores are 0.02. I thought once the correct mAP is computed the LB score should be fine. This is the first time I enter a detection competition and couldn't figure out why. Any suggestions would be appreciated. Thanks!\n\nPS: The mAP calculation code is borrowed from public kernels.",
    "707921": "Sounds like you are overfitting to the training data. What size split are you using for local validation?",
    "708298": "Thanks for your reply. The problem is solved. I'm not overfitting the training data but using the regression `x, y, z` values as submitting results while using the coordinates from the output heatmap to calculate local mAP. However, the 3d regression `x, y, z` doesn't always match the 2d coordinates and this leads to large gap between local mAP and LB score."
  },
  "source": "meta"
}