{
  "id": 78453,
  "title": "Training Images Cropped & Masked",
  "url": "/competitions/humpback-whale-identification/discussion/78453",
  "author_name": "",
  "post_date": "2019-01-24T00:13:00.460757300Z",
  "votes": 69,
  "comment_count": 21,
  "views": 0,
  "content": "<p>Attached are all of the training images cropped and masked. Most of the masks seem pretty good, some of the masks (like the last one) could be improved, to say the least.</p>\n\n<p><img src=\"https://i.imgur.com/mijRPWZ.jpg\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://i.imgur.com/99WkQAm.jpg\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://i.imgur.com/zKk8BuG.jpg\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://i.imgur.com/NhD88KO.jpg\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://i.imgur.com/BnYA2sc.jpg\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://i.imgur.com/B7cGpFw.jpg\" alt=\"enter image description here\"></p>",
  "messages": [
    {
      "id": "460550",
      "postDate": "01/24/2019 00:13:00",
      "content": "<p>Attached are all of the training images cropped and masked. Most of the masks seem pretty good, some of the masks (like the last one) could be improved, to say the least.</p>\n\n<p><img src=\"https://i.imgur.com/mijRPWZ.jpg\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://i.imgur.com/99WkQAm.jpg\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://i.imgur.com/zKk8BuG.jpg\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://i.imgur.com/NhD88KO.jpg\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://i.imgur.com/BnYA2sc.jpg\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://i.imgur.com/B7cGpFw.jpg\" alt=\"enter image description here\"></p>",
      "rawMarkdown": "Attached are all of the training images cropped and masked. Most of the masks seem pretty good, some of the masks (like the last one) could be improved, to say the least.\n\n![enter image description here][1]\n\n![enter image description here][2]\n\n![enter image description here][3]\n\n![enter image description here][4]\n\n![enter image description here][5]\n\n![enter image description here][6]\n\n\n  [1]: https://i.imgur.com/mijRPWZ.jpg\n  [2]: https://i.imgur.com/99WkQAm.jpg\n  [3]: https://i.imgur.com/zKk8BuG.jpg\n  [4]: https://i.imgur.com/NhD88KO.jpg\n  [5]: https://i.imgur.com/BnYA2sc.jpg\n  [6]: https://i.imgur.com/B7cGpFw.jpg",
      "votes": null
    },
    {
      "id": "460912",
      "postDate": "01/24/2019 17:26:55",
      "content": "<p>Great job and thank you for sharing!</p>\n\n<p>What did the trick?</p>",
      "rawMarkdown": "Great job and thank you for sharing!\n\nWhat did the trick?",
      "votes": null
    },
    {
      "id": "461002",
      "postDate": "01/25/2019 01:10:52",
      "content": "<p>First I labeled the foreground/background points on a few hundred images.</p>\n\n<p><img src=\"https://i.imgur.com/5sZ3agl.png?1\" alt=\"'enter description\"></p>\n\n<p>Then after that I used a watershedding algorithm to create masks for the images I labeled. Once I had some masks I trained a U-Net model and then used that to predict masks on the rest of the images.</p>",
      "rawMarkdown": "First I labeled the foreground/background points on a few hundred images.\n\n!['enter description][1]\n\nThen after that I used a watershedding algorithm to create masks for the images I labeled. Once I had some masks I trained a U-Net model and then used that to predict masks on the rest of the images.\n\n[1]: https://i.imgur.com/5sZ3agl.png?1",
      "votes": null
    },
    {
      "id": "461004",
      "postDate": "01/25/2019 01:14:46",
      "content": "<p>Thank you very much!</p>",
      "rawMarkdown": "Thank you very much!",
      "votes": null
    },
    {
      "id": "461015",
      "postDate": "01/25/2019 01:43:39",
      "content": "<p>Thanks Branden!</p>",
      "rawMarkdown": "Thanks Branden!",
      "votes": null
    },
    {
      "id": "461313",
      "postDate": "01/25/2019 18:36:12",
      "content": "<p>Hi Branden and thanks for the annotations! Could you please also provide the annotations for the test data?</p>",
      "rawMarkdown": "Hi Branden and thanks for the annotations! Could you please also provide the annotations for the test data?",
      "votes": null
    },
    {
      "id": "461516",
      "postDate": "01/26/2019 08:36:18",
      "content": "<p>Thank you Branden. Just a small question as I am new to kaggle challenges. Am I allowed to use these for training purpose. ??</p>",
      "rawMarkdown": "Thank you Branden. Just a small question as I am new to kaggle challenges. Am I allowed to use these for training purpose. ??",
      "votes": null
    },
    {
      "id": "461554",
      "postDate": "01/26/2019 10:40:32",
      "content": "<p>Very useful. Thanks Branden</p>",
      "rawMarkdown": "Very useful. Thanks Branden",
      "votes": null
    },
    {
      "id": "461580",
      "postDate": "01/26/2019 12:29:26",
      "content": "<p>Do you have confidence estimates, so we can remove those images which are badly segmented?</p>",
      "rawMarkdown": "Do you have confidence estimates, so we can remove those images which are badly segmented?",
      "votes": null
    },
    {
      "id": "461598",
      "postDate": "01/26/2019 13:17:24",
      "content": "<p>Yes, you are, as it is shared with everyone here.</p>",
      "rawMarkdown": "Yes, you are, as it is shared with everyone here.",
      "votes": null
    },
    {
      "id": "461853",
      "postDate": "01/27/2019 06:56:01",
      "content": "<p>ty</p>",
      "rawMarkdown": "ty",
      "votes": null
    },
    {
      "id": "462044",
      "postDate": "01/27/2019 14:42:16",
      "content": "<p>That is really cool! Thx for the explanation. Could you please also share what program did you use to annotate the images?</p>",
      "rawMarkdown": "That is really cool! Thx for the explanation. Could you please also share what program did you use to annotate the images?",
      "votes": null
    },
    {
      "id": "462111",
      "postDate": "01/27/2019 17:50:00",
      "content": "<p>Great job, Branden! Thank you!</p>",
      "rawMarkdown": "Great job, Branden! Thank you!",
      "votes": null
    },
    {
      "id": "462228",
      "postDate": "01/27/2019 22:55:40",
      "content": "<p>Hi Branden,\nRe <code>I trained a U-Net model and then used that to predict masks on the rest of the images</code>\nI know it's a lot to ask for, but as you decided to share this work maybe you as well can share the link to U-net you used or consider adding it as a kernel. It would be nice to process test in the same way. For those of us with no experience in segmentation, it would be helpful.</p>",
      "rawMarkdown": "Hi Branden,\nRe ```I trained a U-Net model and then used that to predict masks on the rest of the images```\nI know it's a lot to ask for, but as you decided to share this work maybe you as well can share the link to U-net you used or consider adding it as a kernel. It would be nice to process test in the same way. For those of us with no experience in segmentation, it would be helpful.",
      "votes": null
    },
    {
      "id": "462234",
      "postDate": "01/27/2019 23:28:21",
      "content": "<p>great ! Thank you.</p>",
      "rawMarkdown": "great ! Thank you.",
      "votes": null
    },
    {
      "id": "462238",
      "postDate": "01/27/2019 23:49:32",
      "content": "<p>That is interesting technique, thanks a lot! Is it less time-consuming, than trying to label the boundary?</p>",
      "rawMarkdown": "That is interesting technique, thanks a lot! Is it less time-consuming, than trying to label the boundary?",
      "votes": null
    },
    {
      "id": "462360",
      "postDate": "01/28/2019 07:24:34",
      "content": "<p>Nope, the implementation I'm using doesn't output confidence estimates.</p>",
      "rawMarkdown": "Nope, the implementation I'm using doesn't output confidence estimates.",
      "votes": null
    },
    {
      "id": "462366",
      "postDate": "01/28/2019 07:36:12",
      "content": "<p>@radek I use <code>sloth</code>.</p>\n\n<p>@Blonde I use the implementation here: <a href=\"https://github.com/petrosgk/Kaggle-Carvana-Image-Masking-Challenge\">https://github.com/petrosgk/Kaggle-Carvana-Image-Masking-Challenge</a></p>\n\n<p>@old-ufo It would be, but after a while <code>sloth</code> starts to lag behind where I'm clicking so I have to wait for it to catch up. I think it's faster, but not by much and I can do other things while I wait for it to catch up. Maybe there's another annotation tool out there that doesn't lag behind so much. A few of the other annotation tools I tried had even worse issues to deal with, e.g. having to select a \"class\" after creating each point, making the process even longer.</p>",
      "rawMarkdown": "radek I use `sloth`.\n\n@Blonde I use the implementation here: https://github.com/petrosgk/Kaggle-Carvana-Image-Masking-Challenge\n\n@old-ufo It would be, but after a while `sloth` starts to lag behind where I'm clicking so I have to wait for it to catch up. I think it's faster, but not by much and I can do other things while I wait for it to catch up. Maybe there's another annotation tool out there that doesn't lag behind so much. A few of the other annotation tools I tried had even worse issues to deal with, e.g. having to select a \"class\" after creating each point, making the process even longer.",
      "votes": null
    },
    {
      "id": "463462",
      "postDate": "01/30/2019 03:04:48",
      "content": "<p>Hi Branden, or else, could you please share the ground truth you labelled in order to help us to decide which images can be used for validation ? </p>",
      "rawMarkdown": "Hi Branden, or else, could you please share the ground truth you labelled in order to help us to decide which images can be used for validation ?",
      "votes": null
    },
    {
      "id": "469269",
      "postDate": "02/10/2019 20:37:30",
      "content": "<p><a href=\"/vshakhray\">@vshakhray</a> Here are the test images</p>",
      "rawMarkdown": "vshakhray Here are the test images",
      "votes": null
    },
    {
      "id": "471775",
      "postDate": "02/14/2019 21:44:36",
      "content": "<p>Hi Branden  Thank you very much!!</p>\n\n<p>May I ask if you have and could share the cropped but un-masked images.</p>\n\n<p>Thanks!</p>",
      "rawMarkdown": "Hi Branden  Thank you very much!!\n\nMay I ask if you have and could share the cropped but un-masked images.\n\nThanks!",
      "votes": null
    },
    {
      "id": "478516",
      "postDate": "02/26/2019 08:57:18",
      "content": "<p>This is great work Branden ! Thanks for sharing !</p>",
      "rawMarkdown": "This is great work Branden ! Thanks for sharing !",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 460912,
      "author_name": "jeffkk",
      "author_url": "",
      "post_date": "01/24/2019 17:26:55",
      "content": "<p>Great job and thank you for sharing!</p>\n\n<p>What did the trick?</p>",
      "votes": null,
      "replies": [
        {
          "id": 461002,
          "author_name": "brandenkmurray",
          "author_url": "",
          "post_date": "01/25/2019 01:10:52",
          "content": "<p>First I labeled the foreground/background points on a few hundred images.</p>\n\n<p><img src=\"https://i.imgur.com/5sZ3agl.png?1\" alt=\"'enter description\"></p>\n\n<p>Then after that I used a watershedding algorithm to create masks for the images I labeled. Once I had some masks I trained a U-Net model and then used that to predict masks on the rest of the images.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 462044,
          "author_name": "radek1",
          "author_url": "",
          "post_date": "01/27/2019 14:42:16",
          "content": "<p>That is really cool! Thx for the explanation. Could you please also share what program did you use to annotate the images?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 462228,
          "author_name": "blondinka",
          "author_url": "",
          "post_date": "01/27/2019 22:55:40",
          "content": "<p>Hi Branden,\nRe <code>I trained a U-Net model and then used that to predict masks on the rest of the images</code>\nI know it's a lot to ask for, but as you decided to share this work maybe you as well can share the link to U-net you used or consider adding it as a kernel. It would be nice to process test in the same way. For those of us with no experience in segmentation, it would be helpful.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 462238,
          "author_name": "oldufo",
          "author_url": "",
          "post_date": "01/27/2019 23:49:32",
          "content": "<p>That is interesting technique, thanks a lot! Is it less time-consuming, than trying to label the boundary?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 462366,
          "author_name": "brandenkmurray",
          "author_url": "",
          "post_date": "01/28/2019 07:36:12",
          "content": "<p>@radek I use <code>sloth</code>.</p>\n\n<p>@Blonde I use the implementation here: <a href=\"https://github.com/petrosgk/Kaggle-Carvana-Image-Masking-Challenge\">https://github.com/petrosgk/Kaggle-Carvana-Image-Masking-Challenge</a></p>\n\n<p>@old-ufo It would be, but after a while <code>sloth</code> starts to lag behind where I'm clicking so I have to wait for it to catch up. I think it's faster, but not by much and I can do other things while I wait for it to catch up. Maybe there's another annotation tool out there that doesn't lag behind so much. A few of the other annotation tools I tried had even worse issues to deal with, e.g. having to select a \"class\" after creating each point, making the process even longer.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 461004,
      "author_name": "alexanderliao",
      "author_url": "",
      "post_date": "01/25/2019 01:14:46",
      "content": "<p>Thank you very much!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 461015,
      "author_name": "yiheng",
      "author_url": "",
      "post_date": "01/25/2019 01:43:39",
      "content": "<p>Thanks Branden!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 461313,
      "author_name": "vshakhray",
      "author_url": "",
      "post_date": "01/25/2019 18:36:12",
      "content": "<p>Hi Branden and thanks for the annotations! Could you please also provide the annotations for the test data?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 461516,
      "author_name": "tolaniht01",
      "author_url": "",
      "post_date": "01/26/2019 08:36:18",
      "content": "<p>Thank you Branden. Just a small question as I am new to kaggle challenges. Am I allowed to use these for training purpose. ??</p>",
      "votes": null,
      "replies": [
        {
          "id": 461598,
          "author_name": "vshakhray",
          "author_url": "",
          "post_date": "01/26/2019 13:17:24",
          "content": "<p>Yes, you are, as it is shared with everyone here.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 461554,
      "author_name": "lemonic",
      "author_url": "",
      "post_date": "01/26/2019 10:40:32",
      "content": "<p>Very useful. Thanks Branden</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 461580,
      "author_name": "asanakoev",
      "author_url": "",
      "post_date": "01/26/2019 12:29:26",
      "content": "<p>Do you have confidence estimates, so we can remove those images which are badly segmented?</p>",
      "votes": null,
      "replies": [
        {
          "id": 462360,
          "author_name": "brandenkmurray",
          "author_url": "",
          "post_date": "01/28/2019 07:24:34",
          "content": "<p>Nope, the implementation I'm using doesn't output confidence estimates.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 463462,
          "author_name": "yiheng",
          "author_url": "",
          "post_date": "01/30/2019 03:04:48",
          "content": "<p>Hi Branden, or else, could you please share the ground truth you labelled in order to help us to decide which images can be used for validation ? </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 461853,
      "author_name": "darkside6lues",
      "author_url": "",
      "post_date": "01/27/2019 06:56:01",
      "content": "<p>ty</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 462111,
      "author_name": "bejeweled",
      "author_url": "",
      "post_date": "01/27/2019 17:50:00",
      "content": "<p>Great job, Branden! Thank you!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 462234,
      "author_name": "fabsod",
      "author_url": "",
      "post_date": "01/27/2019 23:28:21",
      "content": "<p>great ! Thank you.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 469269,
      "author_name": "brandenkmurray",
      "author_url": "",
      "post_date": "02/10/2019 20:37:30",
      "content": "<p><a href=\"/vshakhray\">@vshakhray</a> Here are the test images</p>",
      "votes": null,
      "replies": [
        {
          "id": 471775,
          "author_name": "yl1202",
          "author_url": "",
          "post_date": "02/14/2019 21:44:36",
          "content": "<p>Hi Branden  Thank you very much!!</p>\n\n<p>May I ask if you have and could share the cropped but un-masked images.</p>\n\n<p>Thanks!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 478516,
      "author_name": "stephanecouvreur",
      "author_url": "",
      "post_date": "02/26/2019 08:57:18",
      "content": "<p>This is great work Branden ! Thanks for sharing !</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "460550": "Attached are all of the training images cropped and masked. Most of the masks seem pretty good, some of the masks (like the last one) could be improved, to say the least.\n\n![enter image description here][1]\n\n![enter image description here][2]\n\n![enter image description here][3]\n\n![enter image description here][4]\n\n![enter image description here][5]\n\n![enter image description here][6]\n\n\n  [1]: https://i.imgur.com/mijRPWZ.jpg\n  [2]: https://i.imgur.com/99WkQAm.jpg\n  [3]: https://i.imgur.com/zKk8BuG.jpg\n  [4]: https://i.imgur.com/NhD88KO.jpg\n  [5]: https://i.imgur.com/BnYA2sc.jpg\n  [6]: https://i.imgur.com/B7cGpFw.jpg",
    "460912": "Great job and thank you for sharing!\n\nWhat did the trick?",
    "461002": "First I labeled the foreground/background points on a few hundred images.\n\n!['enter description][1]\n\nThen after that I used a watershedding algorithm to create masks for the images I labeled. Once I had some masks I trained a U-Net model and then used that to predict masks on the rest of the images.\n\n[1]: https://i.imgur.com/5sZ3agl.png?1",
    "461004": "Thank you very much!",
    "461015": "Thanks Branden!",
    "461313": "Hi Branden and thanks for the annotations! Could you please also provide the annotations for the test data?",
    "461516": "Thank you Branden. Just a small question as I am new to kaggle challenges. Am I allowed to use these for training purpose. ??",
    "461554": "Very useful. Thanks Branden",
    "461580": "Do you have confidence estimates, so we can remove those images which are badly segmented?",
    "461598": "Yes, you are, as it is shared with everyone here.",
    "461853": "ty",
    "462044": "That is really cool! Thx for the explanation. Could you please also share what program did you use to annotate the images?",
    "462111": "Great job, Branden! Thank you!",
    "462228": "Hi Branden,\nRe ```I trained a U-Net model and then used that to predict masks on the rest of the images```\nI know it's a lot to ask for, but as you decided to share this work maybe you as well can share the link to U-net you used or consider adding it as a kernel. It would be nice to process test in the same way. For those of us with no experience in segmentation, it would be helpful.",
    "462234": "great ! Thank you.",
    "462238": "That is interesting technique, thanks a lot! Is it less time-consuming, than trying to label the boundary?",
    "462360": "Nope, the implementation I'm using doesn't output confidence estimates.",
    "462366": "radek I use `sloth`.\n\n@Blonde I use the implementation here: https://github.com/petrosgk/Kaggle-Carvana-Image-Masking-Challenge\n\n@old-ufo It would be, but after a while `sloth` starts to lag behind where I'm clicking so I have to wait for it to catch up. I think it's faster, but not by much and I can do other things while I wait for it to catch up. Maybe there's another annotation tool out there that doesn't lag behind so much. A few of the other annotation tools I tried had even worse issues to deal with, e.g. having to select a \"class\" after creating each point, making the process even longer.",
    "463462": "Hi Branden, or else, could you please share the ground truth you labelled in order to help us to decide which images can be used for validation ?",
    "469269": "vshakhray Here are the test images",
    "471775": "Hi Branden  Thank you very much!!\n\nMay I ask if you have and could share the cropped but un-masked images.\n\nThanks!",
    "478516": "This is great work Branden ! Thanks for sharing !"
  },
  "source": "meta"
}