{
  "id": 262289,
  "title": "LB榜上 mAP 虚高的其中一个原因",
  "url": "/competitions/cowboyoutfits/discussion/262289",
  "author_name": "",
  "post_date": "2021-08-06T09:04:23.407402600Z",
  "votes": 1,
  "comment_count": 4,
  "views": 0,
  "content": "<p>今天调了下模型参数，重新提交了，但是 LB 算出来的 score 很奇怪没有任何变化，目前原因未知，但debug 过程中，似乎找到了 PB 虚高的其中一个原因。</p>\n<p>-- 这是从PB上下载的，评分详细信息，重点关注 【area= small； area=medium 】这两行<br>\n** Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.590**<br>\n Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.947<br>\n Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.456<br>\n Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000<br>\n Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = -1.000<br>\n** Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.590**<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.480<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.637<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.637<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = -1.000<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.637</p>\n<p>可以看出，valid 数据lable中，根本没有中小物体信息，最终 AP 直接等于 【area= large】</p>\n<p>而在我自己验证集上，比如这一份：【area= small； area=medium 】显著低于 large，<br>\n最终 MAP 也低于 large</p>\n<p>Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.317<br>\n Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.474<br>\n Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.348<br>\n Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.129<br>\n Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.269<br>\n Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.339<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.320<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.493<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.500<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.234<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.471<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.528</p>\n<p>但是自己验证集上，large 的 AP 也没有 PB 上那么高，但至少 PB 的验证集的 label 中，没有中小物体，还是非常奇怪的</p>",
  "messages": [
    {
      "id": "1454603",
      "postDate": "08/06/2021 09:04:23",
      "content": "<p>今天调了下模型参数，重新提交了，但是 LB 算出来的 score 很奇怪没有任何变化，目前原因未知，但debug 过程中，似乎找到了 PB 虚高的其中一个原因。</p>\n<p>-- 这是从PB上下载的，评分详细信息，重点关注 【area= small； area=medium 】这两行<br>\n** Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.590**<br>\n Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.947<br>\n Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.456<br>\n Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000<br>\n Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = -1.000<br>\n** Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.590**<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.480<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.637<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.637<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = -1.000<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.637</p>\n<p>可以看出，valid 数据lable中，根本没有中小物体信息，最终 AP 直接等于 【area= large】</p>\n<p>而在我自己验证集上，比如这一份：【area= small； area=medium 】显著低于 large，<br>\n最终 MAP 也低于 large</p>\n<p>Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.317<br>\n Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.474<br>\n Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.348<br>\n Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.129<br>\n Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.269<br>\n Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.339<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.320<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.493<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.500<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.234<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.471<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.528</p>\n<p>但是自己验证集上，large 的 AP 也没有 PB 上那么高，但至少 PB 的验证集的 label 中，没有中小物体，还是非常奇怪的</p>",
      "rawMarkdown": "今天调了下模型参数，重新提交了，但是 LB 算出来的 score 很奇怪没有任何变化，目前原因未知，但debug 过程中，似乎找到了 PB 虚高的其中一个原因。\n\n-- 这是从PB上下载的，评分详细信息，重点关注 【area= small； area=medium 】这两行\n** Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.590**\n Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.947\n Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.456\n Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = -1.000\n** Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.590**\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.480\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.637\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.637\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = -1.000\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.637\n\n可以看出，valid 数据lable中，根本没有中小物体信息，最终 AP 直接等于 【area= large】\n\n而在我自己验证集上，比如这一份：【area= small； area=medium 】显著低于 large，\n最终 MAP 也低于 large\n\n Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.317\n Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.474\n Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.348\n Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.129\n Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.269\n Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.339\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.320\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.493\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.500\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.234\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.471\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.528\n\n但是自己验证集上，large 的 AP 也没有 PB 上那么高，但至少 PB 的验证集的 label 中，没有中小物体，还是非常奇怪的",
      "votes": null
    },
    {
      "id": "1455610",
      "postDate": "08/06/2021 16:13:46",
      "content": "<p>总觉得是哪有问题，我把belt的预测单独提交，得到是0，自己在验证集跑的结果还可以。然后发现image_id会影响这个结果，不知道评估的code里面会不会有bug</p>",
      "rawMarkdown": "总觉得是哪有问题，我把belt的预测单独提交，得到是0，自己在验证集跑的结果还可以。然后发现image_id会影响这个结果，不知道评估的code里面会不会有bug",
      "votes": null
    },
    {
      "id": "1456698",
      "postDate": "08/07/2021 04:00:55",
      "content": "<p>今天才注意到这个，DevPhase只看large，FinalPhase只看medium和large</p>",
      "rawMarkdown": "今天才注意到这个，DevPhase只看large，FinalPhase只看medium和large",
      "votes": null
    },
    {
      "id": "1456917",
      "postDate": "08/07/2021 06:35:52",
      "content": "<p>是的，我又细看了下 final 阶段 score 的结果，而且还有一点，medium 竟然显著高于 large。。。<br>\n无论是我自己的验证集，还是 coco 上各模型榜单，AP 明显 large &gt;&gt; medium &gt;&gt; small 的</p>\n<p>--  resnet50-faster-rcnn-epoch15 结果 -- <br>\ncreating index…<br>\nindex created!<br>\nRunning per image evaluation…<br>\nEvaluate annotation type <em>bbox</em><br>\nDONE (t=0.04s).<br>\nAccumulating evaluation results…<br>\nDONE (t=0.04s).<br>\n Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.460<br>\n Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.694<br>\n Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.493<br>\n Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000<br>\n Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.550<br>\n Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.492<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.469<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.531<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.531<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.550<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.531<br>\nmAP: 46.000707213578494</p>",
      "rawMarkdown": "是的，我又细看了下 final 阶段 score 的结果，而且还有一点，medium 竟然显著高于 large。。。\n无论是我自己的验证集，还是 coco 上各模型榜单，AP 明显 large >> medium >> small 的\n\n--  resnet50-faster-rcnn-epoch15 结果 -- \ncreating index...\nindex created!\nRunning per image evaluation...\nEvaluate annotation type *bbox*\nDONE (t=0.04s).\nAccumulating evaluation results...\nDONE (t=0.04s).\n Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.460\n Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.694\n Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.493\n Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.550\n Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.492\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.469\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.531\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.531\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.550\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.531\nmAP: 46.000707213578494",
      "votes": null
    },
    {
      "id": "1458614",
      "postDate": "08/08/2021 01:23:09",
      "content": "<p>很奇怪你medium和large的mAP都比最后总都要高。我刚看了我最后3个submissions，都是large&gt;final&gt;medium。但奇怪的是有一个的large 和medium都更高，final却更低。一直没搞懂…，猜测有些objects甚至不在这三个范围里面？</p>",
      "rawMarkdown": "很奇怪你medium和large的mAP都比最后总都要高。我刚看了我最后3个submissions，都是large>final>medium。但奇怪的是有一个的large 和medium都更高，final却更低。一直没搞懂...，猜测有些objects甚至不在这三个范围里面？",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1455610,
      "author_name": "yichaohan",
      "author_url": "",
      "post_date": "08/06/2021 16:13:46",
      "content": "<p>总觉得是哪有问题，我把belt的预测单独提交，得到是0，自己在验证集跑的结果还可以。然后发现image_id会影响这个结果，不知道评估的code里面会不会有bug</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1456698,
      "author_name": "snowclem",
      "author_url": "",
      "post_date": "08/07/2021 04:00:55",
      "content": "<p>今天才注意到这个，DevPhase只看large，FinalPhase只看medium和large</p>",
      "votes": null,
      "replies": [
        {
          "id": 1456917,
          "author_name": "niangaopx",
          "author_url": "",
          "post_date": "08/07/2021 06:35:52",
          "content": "<p>是的，我又细看了下 final 阶段 score 的结果，而且还有一点，medium 竟然显著高于 large。。。<br>\n无论是我自己的验证集，还是 coco 上各模型榜单，AP 明显 large &gt;&gt; medium &gt;&gt; small 的</p>\n<p>--  resnet50-faster-rcnn-epoch15 结果 -- <br>\ncreating index…<br>\nindex created!<br>\nRunning per image evaluation…<br>\nEvaluate annotation type <em>bbox</em><br>\nDONE (t=0.04s).<br>\nAccumulating evaluation results…<br>\nDONE (t=0.04s).<br>\n Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.460<br>\n Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.694<br>\n Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.493<br>\n Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000<br>\n Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.550<br>\n Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.492<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.469<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.531<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.531<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.550<br>\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.531<br>\nmAP: 46.000707213578494</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1458614,
          "author_name": "yichaohan",
          "author_url": "",
          "post_date": "08/08/2021 01:23:09",
          "content": "<p>很奇怪你medium和large的mAP都比最后总都要高。我刚看了我最后3个submissions，都是large&gt;final&gt;medium。但奇怪的是有一个的large 和medium都更高，final却更低。一直没搞懂…，猜测有些objects甚至不在这三个范围里面？</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1454603": "今天调了下模型参数，重新提交了，但是 LB 算出来的 score 很奇怪没有任何变化，目前原因未知，但debug 过程中，似乎找到了 PB 虚高的其中一个原因。\n\n-- 这是从PB上下载的，评分详细信息，重点关注 【area= small； area=medium 】这两行\n** Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.590**\n Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.947\n Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.456\n Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = -1.000\n** Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.590**\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.480\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.637\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.637\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = -1.000\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.637\n\n可以看出，valid 数据lable中，根本没有中小物体信息，最终 AP 直接等于 【area= large】\n\n而在我自己验证集上，比如这一份：【area= small； area=medium 】显著低于 large，\n最终 MAP 也低于 large\n\n Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.317\n Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.474\n Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.348\n Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.129\n Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.269\n Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.339\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.320\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.493\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.500\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.234\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.471\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.528\n\n但是自己验证集上，large 的 AP 也没有 PB 上那么高，但至少 PB 的验证集的 label 中，没有中小物体，还是非常奇怪的",
    "1455610": "总觉得是哪有问题，我把belt的预测单独提交，得到是0，自己在验证集跑的结果还可以。然后发现image_id会影响这个结果，不知道评估的code里面会不会有bug",
    "1456698": "今天才注意到这个，DevPhase只看large，FinalPhase只看medium和large",
    "1456917": "是的，我又细看了下 final 阶段 score 的结果，而且还有一点，medium 竟然显著高于 large。。。\n无论是我自己的验证集，还是 coco 上各模型榜单，AP 明显 large >> medium >> small 的\n\n--  resnet50-faster-rcnn-epoch15 结果 -- \ncreating index...\nindex created!\nRunning per image evaluation...\nEvaluate annotation type *bbox*\nDONE (t=0.04s).\nAccumulating evaluation results...\nDONE (t=0.04s).\n Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.460\n Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.694\n Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.493\n Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.550\n Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.492\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.469\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.531\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.531\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.550\n Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.531\nmAP: 46.000707213578494",
    "1458614": "很奇怪你medium和large的mAP都比最后总都要高。我刚看了我最后3个submissions，都是large>final>medium。但奇怪的是有一个的large 和medium都更高，final却更低。一直没搞懂...，猜测有些objects甚至不在这三个范围里面？"
  },
  "source": "meta"
}