{
  "id": 91505,
  "title": "Sharing your local segmentation statistics (w.o. fine-grained classification)",
  "url": "/competitions/imaterialist-fashion-2019-FGVC6/discussion/91505",
  "author_name": "",
  "post_date": "2019-05-05T19:54:38.585538300Z",
  "votes": 2,
  "comment_count": 11,
  "views": 0,
  "content": "<p>As of 5/9/19:</p>\n\n<p>Average Precision@IoU=0.50:0.95 -&gt; 0.287</p>",
  "messages": [
    {
      "id": "527556",
      "postDate": "05/05/2019 19:54:38",
      "content": "<p>As of 5/9/19:</p>\n\n<p>Average Precision@IoU=0.50:0.95 -&gt; 0.287</p>",
      "rawMarkdown": "As of 5/9/19:\n\nAverage Precision@IoU=0.50:0.95 -&gt; 0.287",
      "votes": null
    },
    {
      "id": "527630",
      "postDate": "05/06/2019 01:36:07",
      "content": "<p>Thank you for your sharing！Could you tell me the size of the training? </p>",
      "rawMarkdown": "Thank you for your sharing！Could you tell me the size of the training?",
      "votes": null
    },
    {
      "id": "527632",
      "postDate": "05/06/2019 01:49:17",
      "content": "<p>Do you mean input image size?</p>",
      "rawMarkdown": "Do you mean input image size?",
      "votes": null
    },
    {
      "id": "527741",
      "postDate": "05/06/2019 07:57:39",
      "content": "<p>Here are some validation metrics I have:\n<code>\n                       AP_0.50: 0.2449\n                       AP_0.55: 0.2340\n                       AP_0.60: 0.2228\n                       AP_0.65: 0.2083\n                       AP_0.70: 0.1899\n                       AP_0.75: 0.1661\n                       AP_0.80: 0.1355\n                       AP_0.85: 0.0929\n                       AP_0.90: 0.0421\n                       AP_0.95: 0.0021\n           category_perfect_f1: 0.6631\n           class_id_perfect_f1: 0.4131\nclass_id_strip_attr_perfect_f1: 0.4104\n                           mAP: 0.1539\n</code>\n(\"perfect\" f1 metrics are calculated by matching class_id or category predictions within one image, not taking masks into account).</p>",
      "rawMarkdown": "Here are some validation metrics I have:\n```\n                       AP_0.50: 0.2449\n                       AP_0.55: 0.2340\n                       AP_0.60: 0.2228\n                       AP_0.65: 0.2083\n                       AP_0.70: 0.1899\n                       AP_0.75: 0.1661\n                       AP_0.80: 0.1355\n                       AP_0.85: 0.0929\n                       AP_0.90: 0.0421\n                       AP_0.95: 0.0021\n           category_perfect_f1: 0.6631\n           class_id_perfect_f1: 0.4131\nclass_id_strip_attr_perfect_f1: 0.4104\n                           mAP: 0.1539\n```\n(\"perfect\" f1 metrics are calculated by matching class_id or category predictions within one image, not taking masks into account).",
      "votes": null
    },
    {
      "id": "527849",
      "postDate": "05/06/2019 13:13:06",
      "content": "<p>Yes, I turned the image into 512*512 to train, but I think this is not good for segmentation.</p>",
      "rawMarkdown": "Yes, I turned the image into 512*512 to train, but I think this is not good for segmentation.",
      "votes": null
    },
    {
      "id": "532039",
      "postDate": "05/16/2019 04:22:07",
      "content": "<p>Hey man, I am wondering if your mAP was calculated by taking attributes into consideration?</p>",
      "rawMarkdown": "Hey man, I am wondering if your mAP was calculated by taking attributes into consideration?",
      "votes": null
    },
    {
      "id": "532115",
      "postDate": "05/16/2019 08:18:40",
      "content": "<p><a href=\"/ascust\">@ascust</a> yes, this was calculated taking attributes into account</p>",
      "rawMarkdown": "ascust yes, this was calculated taking attributes into account",
      "votes": null
    },
    {
      "id": "532122",
      "postDate": "05/16/2019 08:36:29",
      "content": "<p>Thanks for your reply. I am just curious how well the attributes contribute to the final score. Since the attributes are required to be exactly same with the ground-truth for each instance, it seems a big challenge.  In other words, how much score did you get (if you have ever tried) if you did not predict any attributes.</p>",
      "rawMarkdown": "Thanks for your reply. I am just curious how well the attributes contribute to the final score. Since the attributes are required to be exactly same with the ground-truth for each instance, it seems a big challenge.  In other words, how much score did you get (if you have ever tried) if you did not predict any attributes.",
      "votes": null
    },
    {
      "id": "532134",
      "postDate": "05/16/2019 09:07:31",
      "content": "<p>Their contribution is extremely minor, attributes prediction is very bad for me.</p>",
      "rawMarkdown": "Their contribution is extremely minor, attributes prediction is very bad for me.",
      "votes": null
    },
    {
      "id": "537609",
      "postDate": "05/27/2019 10:16:19",
      "content": "<p><a href=\"/lopuhin\">@lopuhin</a> any hint as to how we can predict the attribute ids? The only thing I can think of now is to train another mask rcnn with one hot encoded output? </p>",
      "rawMarkdown": "lopuhin any hint as to how we can predict the attribute ids? The only thing I can think of now is to train another mask rcnn with one hot encoded output?",
      "votes": null
    },
    {
      "id": "537673",
      "postDate": "05/27/2019 12:50:02",
      "content": "<p>We have a branch in mask-rcnn which predicts attributes along with labels, but it's working extremely poorly. And we also experimented with separate image classification models. Not sure how to best deal with attributes myself.</p>",
      "rawMarkdown": "We have a branch in mask-rcnn which predicts attributes along with labels, but it's working extremely poorly. And we also experimented with separate image classification models. Not sure how to best deal with attributes myself.",
      "votes": null
    },
    {
      "id": "538872",
      "postDate": "05/29/2019 08:15:05",
      "content": "<p>And any tips on how to improve the performance? Like everyone uses the mask-RCNN, but what makes a good score?</p>",
      "rawMarkdown": "And any tips on how to improve the performance? Like everyone uses the mask-RCNN, but what makes a good score?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 527630,
      "author_name": "hesene",
      "author_url": "",
      "post_date": "05/06/2019 01:36:07",
      "content": "<p>Thank you for your sharing！Could you tell me the size of the training? </p>",
      "votes": null,
      "replies": [
        {
          "id": 527632,
          "author_name": "alexanderliao",
          "author_url": "",
          "post_date": "05/06/2019 01:49:17",
          "content": "<p>Do you mean input image size?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 527849,
          "author_name": "hesene",
          "author_url": "",
          "post_date": "05/06/2019 13:13:06",
          "content": "<p>Yes, I turned the image into 512*512 to train, but I think this is not good for segmentation.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 527741,
      "author_name": "lopuhin",
      "author_url": "",
      "post_date": "05/06/2019 07:57:39",
      "content": "<p>Here are some validation metrics I have:\n<code>\n                       AP_0.50: 0.2449\n                       AP_0.55: 0.2340\n                       AP_0.60: 0.2228\n                       AP_0.65: 0.2083\n                       AP_0.70: 0.1899\n                       AP_0.75: 0.1661\n                       AP_0.80: 0.1355\n                       AP_0.85: 0.0929\n                       AP_0.90: 0.0421\n                       AP_0.95: 0.0021\n           category_perfect_f1: 0.6631\n           class_id_perfect_f1: 0.4131\nclass_id_strip_attr_perfect_f1: 0.4104\n                           mAP: 0.1539\n</code>\n(\"perfect\" f1 metrics are calculated by matching class_id or category predictions within one image, not taking masks into account).</p>",
      "votes": null,
      "replies": [
        {
          "id": 532039,
          "author_name": "ascust",
          "author_url": "",
          "post_date": "05/16/2019 04:22:07",
          "content": "<p>Hey man, I am wondering if your mAP was calculated by taking attributes into consideration?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 532115,
          "author_name": "lopuhin",
          "author_url": "",
          "post_date": "05/16/2019 08:18:40",
          "content": "<p><a href=\"/ascust\">@ascust</a> yes, this was calculated taking attributes into account</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 532122,
          "author_name": "ascust",
          "author_url": "",
          "post_date": "05/16/2019 08:36:29",
          "content": "<p>Thanks for your reply. I am just curious how well the attributes contribute to the final score. Since the attributes are required to be exactly same with the ground-truth for each instance, it seems a big challenge.  In other words, how much score did you get (if you have ever tried) if you did not predict any attributes.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 532134,
          "author_name": "lopuhin",
          "author_url": "",
          "post_date": "05/16/2019 09:07:31",
          "content": "<p>Their contribution is extremely minor, attributes prediction is very bad for me.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 537609,
      "author_name": "soulshepherd",
      "author_url": "",
      "post_date": "05/27/2019 10:16:19",
      "content": "<p><a href=\"/lopuhin\">@lopuhin</a> any hint as to how we can predict the attribute ids? The only thing I can think of now is to train another mask rcnn with one hot encoded output? </p>",
      "votes": null,
      "replies": [
        {
          "id": 537673,
          "author_name": "lopuhin",
          "author_url": "",
          "post_date": "05/27/2019 12:50:02",
          "content": "<p>We have a branch in mask-rcnn which predicts attributes along with labels, but it's working extremely poorly. And we also experimented with separate image classification models. Not sure how to best deal with attributes myself.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 538872,
          "author_name": "soulshepherd",
          "author_url": "",
          "post_date": "05/29/2019 08:15:05",
          "content": "<p>And any tips on how to improve the performance? Like everyone uses the mask-RCNN, but what makes a good score?</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "527556": "As of 5/9/19:\n\nAverage Precision@IoU=0.50:0.95 -&gt; 0.287",
    "527630": "Thank you for your sharing！Could you tell me the size of the training?",
    "527632": "Do you mean input image size?",
    "527741": "Here are some validation metrics I have:\n```\n                       AP_0.50: 0.2449\n                       AP_0.55: 0.2340\n                       AP_0.60: 0.2228\n                       AP_0.65: 0.2083\n                       AP_0.70: 0.1899\n                       AP_0.75: 0.1661\n                       AP_0.80: 0.1355\n                       AP_0.85: 0.0929\n                       AP_0.90: 0.0421\n                       AP_0.95: 0.0021\n           category_perfect_f1: 0.6631\n           class_id_perfect_f1: 0.4131\nclass_id_strip_attr_perfect_f1: 0.4104\n                           mAP: 0.1539\n```\n(\"perfect\" f1 metrics are calculated by matching class_id or category predictions within one image, not taking masks into account).",
    "527849": "Yes, I turned the image into 512*512 to train, but I think this is not good for segmentation.",
    "532039": "Hey man, I am wondering if your mAP was calculated by taking attributes into consideration?",
    "532115": "ascust yes, this was calculated taking attributes into account",
    "532122": "Thanks for your reply. I am just curious how well the attributes contribute to the final score. Since the attributes are required to be exactly same with the ground-truth for each instance, it seems a big challenge.  In other words, how much score did you get (if you have ever tried) if you did not predict any attributes.",
    "532134": "Their contribution is extremely minor, attributes prediction is very bad for me.",
    "537609": "lopuhin any hint as to how we can predict the attribute ids? The only thing I can think of now is to train another mask rcnn with one hot encoded output?",
    "537673": "We have a branch in mask-rcnn which predicts attributes along with labels, but it's working extremely poorly. And we also experimented with separate image classification models. Not sure how to best deal with attributes myself.",
    "538872": "And any tips on how to improve the performance? Like everyone uses the mask-RCNN, but what makes a good score?"
  },
  "source": "meta"
}