{
  "id": 33253,
  "title": "Bounding boxes - anyone to annotate?",
  "url": "/competitions/noaa-fisheries-steller-sea-lion-population-count/discussion/33253",
  "author_name": "",
  "post_date": "2017-05-19T15:23:07.290034800Z",
  "votes": 1,
  "comment_count": 10,
  "views": 0,
  "content": "<p>Hi all, </p>\n\n<p>I work on object detection using deep CNNs in my everyday work. I am confident the same approach I use at work will work very well here. </p>\n\n<p>Ideally, the training data should have been in the form of bounding boxes (instead of coloured dots). Is anyone willing to team up with me and annotate as many images as possible by the end of May?\nI will be more free by then and can start training CNNs.</p>\n\n<p>I tried Sloth <a href=\"http://sloth.readthedocs.io/\">http://sloth.readthedocs.io</a> and it works fine. It produces annotations like this:</p>\n\n<pre><code>        {\n            \"class\": \"pup\", \n            \"height\": 26.341403930330898, \n            \"type\": \"rect\", \n            \"width\": 16.35939823041599, \n            \"x\": 3256.629359597222, \n            \"y\": 1671.9859547357385\n        }, \n        {\n            \"class\": \"pup\", \n            \"height\": 24.40045837756952, \n            \"type\": \"rect\", \n            \"width\": 22.459512824808826, \n            \"x\": 3230.287955666891, \n            \"y\": 1709.695754046528\n        }, \n        {\n            \"class\": \"pup\", \n            \"height\": 24.283655282917607, \n            \"type\": \"rect\", \n            \"width\": 20.875422962507855, \n            \"x\": 3009.469138921581, \n            \"y\": 1819.9960590986684\n        }, \n        {\n            \"class\": \"female\", \n            \"height\": 88.61404033064673, \n            \"type\": \"rect\", \n            \"width\": 36.2124684043506, \n            \"x\": 2975.8128447575373, \n            \"y\": 1664.0694304399342\n        }, \n        {\n            \"class\": \"female\", \n            \"height\": 43.02893304516988, \n            \"type\": \"rect\", \n            \"width\": 55.809804246705426, \n            \"x\": 3027.788387643782, \n            \"y\": 1645.324152677682\n        }, \n</code></pre>\n\n<p>Looking forward to collaborating. Hit me up, thanks!</p>",
  "messages": [
    {
      "id": "183881",
      "postDate": "05/19/2017 15:23:07",
      "content": "<p>Hi all, </p>\n\n<p>I work on object detection using deep CNNs in my everyday work. I am confident the same approach I use at work will work very well here. </p>\n\n<p>Ideally, the training data should have been in the form of bounding boxes (instead of coloured dots). Is anyone willing to team up with me and annotate as many images as possible by the end of May?\nI will be more free by then and can start training CNNs.</p>\n\n<p>I tried Sloth <a href=\"http://sloth.readthedocs.io/\">http://sloth.readthedocs.io</a> and it works fine. It produces annotations like this:</p>\n\n<pre><code>        {\n            \"class\": \"pup\", \n            \"height\": 26.341403930330898, \n            \"type\": \"rect\", \n            \"width\": 16.35939823041599, \n            \"x\": 3256.629359597222, \n            \"y\": 1671.9859547357385\n        }, \n        {\n            \"class\": \"pup\", \n            \"height\": 24.40045837756952, \n            \"type\": \"rect\", \n            \"width\": 22.459512824808826, \n            \"x\": 3230.287955666891, \n            \"y\": 1709.695754046528\n        }, \n        {\n            \"class\": \"pup\", \n            \"height\": 24.283655282917607, \n            \"type\": \"rect\", \n            \"width\": 20.875422962507855, \n            \"x\": 3009.469138921581, \n            \"y\": 1819.9960590986684\n        }, \n        {\n            \"class\": \"female\", \n            \"height\": 88.61404033064673, \n            \"type\": \"rect\", \n            \"width\": 36.2124684043506, \n            \"x\": 2975.8128447575373, \n            \"y\": 1664.0694304399342\n        }, \n        {\n            \"class\": \"female\", \n            \"height\": 43.02893304516988, \n            \"type\": \"rect\", \n            \"width\": 55.809804246705426, \n            \"x\": 3027.788387643782, \n            \"y\": 1645.324152677682\n        }, \n</code></pre>\n\n<p>Looking forward to collaborating. Hit me up, thanks!</p>",
      "rawMarkdown": "Hi all, \n\nI work on object detection using deep CNNs in my everyday work. I am confident the same approach I use at work will work very well here. \n\nIdeally, the training data should have been in the form of bounding boxes (instead of coloured dots). Is anyone willing to team up with me and annotate as many images as possible by the end of May?\nI will be more free by then and can start training CNNs.\n\nI tried Sloth http://sloth.readthedocs.io and it works fine. It produces annotations like this:\n\n            {\n                \"class\": \"pup\", \n                \"height\": 26.341403930330898, \n                \"type\": \"rect\", \n                \"width\": 16.35939823041599, \n                \"x\": 3256.629359597222, \n                \"y\": 1671.9859547357385\n            }, \n            {\n                \"class\": \"pup\", \n                \"height\": 24.40045837756952, \n                \"type\": \"rect\", \n                \"width\": 22.459512824808826, \n                \"x\": 3230.287955666891, \n                \"y\": 1709.695754046528\n            }, \n            {\n                \"class\": \"pup\", \n                \"height\": 24.283655282917607, \n                \"type\": \"rect\", \n                \"width\": 20.875422962507855, \n                \"x\": 3009.469138921581, \n                \"y\": 1819.9960590986684\n            }, \n            {\n                \"class\": \"female\", \n                \"height\": 88.61404033064673, \n                \"type\": \"rect\", \n                \"width\": 36.2124684043506, \n                \"x\": 2975.8128447575373, \n                \"y\": 1664.0694304399342\n            }, \n            {\n                \"class\": \"female\", \n                \"height\": 43.02893304516988, \n                \"type\": \"rect\", \n                \"width\": 55.809804246705426, \n                \"x\": 3027.788387643782, \n                \"y\": 1645.324152677682\n            }, \n\nLooking forward to collaborating. Hit me up, thanks!",
      "votes": null
    },
    {
      "id": "183882",
      "postDate": "05/19/2017 15:32:41",
      "content": "<p>Hi Chan,\nI already did the bounding boxes annotations using a simple heuristic, considering the same box size per class. It works decent and straight-forward. Those are my delta values from the center:\ndeltas = {'pups': {'x': -12, 'y': 12},\n                       'juveniles': {'x': -40, 'y': 40},\n                       'subadult_males': {'x': -30, 'y': 30},\n                       'adult_females': {'x': -25, 'y': 25},\n                       'adult_males': {'x': -30, 'y': 30}}  </p>",
      "rawMarkdown": "Hi Chan,\nI already did the bounding boxes annotations using a simple heuristic, considering the same box size per class. It works decent and straight-forward. Those are my delta values from the center:\ndeltas = {'pups': {'x': -12, 'y': 12},\n                       'juveniles': {'x': -40, 'y': 40},\n                       'subadult_males': {'x': -30, 'y': 30},\n                       'adult_females': {'x': -25, 'y': 25},\n                       'adult_males': {'x': -30, 'y': 30}}",
      "votes": null
    },
    {
      "id": "183992",
      "postDate": "05/20/2017 00:59:05",
      "content": "<p>Have you looked into using Amazon Mechanical Turk? That could be the way to go with this.</p>",
      "rawMarkdown": "Have you looked into using Amazon Mechanical Turk? That could be the way to go with this.",
      "votes": null
    },
    {
      "id": "183996",
      "postDate": "05/20/2017 01:22:55",
      "content": "<p>I am interested in collaborating. </p>",
      "rawMarkdown": "I am interested in collaborating.",
      "votes": null
    },
    {
      "id": "184379",
      "postDate": "05/21/2017 15:52:49",
      "content": "<p>The problem is the images are taken from different altitudes. In some cases, adult males are only 60 pixels, and in some images, adult male is over 210 pixels end to end.</p>",
      "rawMarkdown": "The problem is the images are taken from different altitudes. In some cases, adult males are only 60 pixels, and in some images, adult male is over 210 pixels end to end.",
      "votes": null
    },
    {
      "id": "184380",
      "postDate": "05/21/2017 15:53:19",
      "content": "<p>I think that'd violate the contest rules.</p>",
      "rawMarkdown": "I think that'd violate the contest rules.",
      "votes": null
    },
    {
      "id": "184741",
      "postDate": "05/22/2017 21:55:17",
      "content": "<p>I am interested</p>",
      "rawMarkdown": "I am interested",
      "votes": null
    },
    {
      "id": "184941",
      "postDate": "05/23/2017 15:50:13",
      "content": "<p>Hi guys, sorry didn't login the last few days. Can you please add me on LinkedIn (contact on my profile) and we can discuss more? I am not able to send messages here.</p>\n\n<p>Also, please try installing and using sloth. It took me a bit of time setting it up. If you manage to do it, I can pass you the config file I wrote for this dataset.</p>",
      "rawMarkdown": "Hi guys, sorry didn't login the last few days. Can you please add me on LinkedIn (contact on my profile) and we can discuss more? I am not able to send messages here.\n\nAlso, please try installing and using sloth. It took me a bit of time setting it up. If you manage to do it, I can pass you the config file I wrote for this dataset.",
      "votes": null
    },
    {
      "id": "186172",
      "postDate": "05/27/2017 00:32:36",
      "content": "<p>I am interested as well</p>",
      "rawMarkdown": "I am interested as well",
      "votes": null
    },
    {
      "id": "186360",
      "postDate": "05/27/2017 19:47:08",
      "content": "<p>I don't think that it matters much. In the end the score regards the number of elements per class not about predicting the right rectangle. An average size can be good enough. </p>",
      "rawMarkdown": "I don't think that it matters much. In the end the score regards the number of elements per class not about predicting the right rectangle. An average size can be good enough.",
      "votes": null
    },
    {
      "id": "193870",
      "postDate": "06/18/2017 14:36:09",
      "content": "<p>If you are using Mac OS X, you can use RectLabel. </p>\n\n<p>An image annotation tool to label images for bounding box object detection and segmentation.</p>\n\n<p><a href=\"https://rectlabel.com/\">https://rectlabel.com</a></p>\n\n<p>Key features:</p>\n\n<ul>\n<li><p>Drawing bounding box, polygon, and cubic bezier</p></li>\n<li><p>Export index color mask image and separated mask images</p></li>\n<li><p>1-click buttons make your labeling work faster</p></li>\n<li><p>Customize the label dialog to combine with attributes</p></li>\n<li><p>Settings for objects, attributes, hotkeys, and labeling fast</p></li>\n<li><p>Layer order for overlapped boxes</p></li>\n<li><p>Quick zoom to existing boxes</p></li>\n<li><p>Support the PASCAL VOC format</p></li>\n</ul>",
      "rawMarkdown": "If you are using Mac OS X, you can use RectLabel. \n\nAn image annotation tool to label images for bounding box object detection and segmentation.\n\nhttps://rectlabel.com\n\nKey features:\n\n- Drawing bounding box, polygon, and cubic bezier\n\n- Export index color mask image and separated mask images\n\n- 1-click buttons make your labeling work faster\n\n- Customize the label dialog to combine with attributes\n\n- Settings for objects, attributes, hotkeys, and labeling fast\n\n- Layer order for overlapped boxes\n\n- Quick zoom to existing boxes\n\n- Support the PASCAL VOC format",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 183882,
      "author_name": "adimar",
      "author_url": "",
      "post_date": "05/19/2017 15:32:41",
      "content": "<p>Hi Chan,\nI already did the bounding boxes annotations using a simple heuristic, considering the same box size per class. It works decent and straight-forward. Those are my delta values from the center:\ndeltas = {'pups': {'x': -12, 'y': 12},\n                       'juveniles': {'x': -40, 'y': 40},\n                       'subadult_males': {'x': -30, 'y': 30},\n                       'adult_females': {'x': -25, 'y': 25},\n                       'adult_males': {'x': -30, 'y': 30}}  </p>",
      "votes": null,
      "replies": [
        {
          "id": 184379,
          "author_name": "authman",
          "author_url": "",
          "post_date": "05/21/2017 15:52:49",
          "content": "<p>The problem is the images are taken from different altitudes. In some cases, adult males are only 60 pixels, and in some images, adult male is over 210 pixels end to end.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 186360,
          "author_name": "adimar",
          "author_url": "",
          "post_date": "05/27/2017 19:47:08",
          "content": "<p>I don't think that it matters much. In the end the score regards the number of elements per class not about predicting the right rectangle. An average size can be good enough. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 183992,
      "author_name": "davidzqhuang",
      "author_url": "",
      "post_date": "05/20/2017 00:59:05",
      "content": "<p>Have you looked into using Amazon Mechanical Turk? That could be the way to go with this.</p>",
      "votes": null,
      "replies": [
        {
          "id": 184380,
          "author_name": "authman",
          "author_url": "",
          "post_date": "05/21/2017 15:53:19",
          "content": "<p>I think that'd violate the contest rules.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 183996,
      "author_name": "walterwangnyc",
      "author_url": "",
      "post_date": "05/20/2017 01:22:55",
      "content": "<p>I am interested in collaborating. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 184741,
      "author_name": "jkarimi91",
      "author_url": "",
      "post_date": "05/22/2017 21:55:17",
      "content": "<p>I am interested</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 184941,
      "author_name": "chandrachud",
      "author_url": "",
      "post_date": "05/23/2017 15:50:13",
      "content": "<p>Hi guys, sorry didn't login the last few days. Can you please add me on LinkedIn (contact on my profile) and we can discuss more? I am not able to send messages here.</p>\n\n<p>Also, please try installing and using sloth. It took me a bit of time setting it up. If you manage to do it, I can pass you the config file I wrote for this dataset.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 186172,
      "author_name": "narras",
      "author_url": "",
      "post_date": "05/27/2017 00:32:36",
      "content": "<p>I am interested as well</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 193870,
      "author_name": "ryouchinsa",
      "author_url": "",
      "post_date": "06/18/2017 14:36:09",
      "content": "<p>If you are using Mac OS X, you can use RectLabel. </p>\n\n<p>An image annotation tool to label images for bounding box object detection and segmentation.</p>\n\n<p><a href=\"https://rectlabel.com/\">https://rectlabel.com</a></p>\n\n<p>Key features:</p>\n\n<ul>\n<li><p>Drawing bounding box, polygon, and cubic bezier</p></li>\n<li><p>Export index color mask image and separated mask images</p></li>\n<li><p>1-click buttons make your labeling work faster</p></li>\n<li><p>Customize the label dialog to combine with attributes</p></li>\n<li><p>Settings for objects, attributes, hotkeys, and labeling fast</p></li>\n<li><p>Layer order for overlapped boxes</p></li>\n<li><p>Quick zoom to existing boxes</p></li>\n<li><p>Support the PASCAL VOC format</p></li>\n</ul>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "183881": "Hi all, \n\nI work on object detection using deep CNNs in my everyday work. I am confident the same approach I use at work will work very well here. \n\nIdeally, the training data should have been in the form of bounding boxes (instead of coloured dots). Is anyone willing to team up with me and annotate as many images as possible by the end of May?\nI will be more free by then and can start training CNNs.\n\nI tried Sloth http://sloth.readthedocs.io and it works fine. It produces annotations like this:\n\n            {\n                \"class\": \"pup\", \n                \"height\": 26.341403930330898, \n                \"type\": \"rect\", \n                \"width\": 16.35939823041599, \n                \"x\": 3256.629359597222, \n                \"y\": 1671.9859547357385\n            }, \n            {\n                \"class\": \"pup\", \n                \"height\": 24.40045837756952, \n                \"type\": \"rect\", \n                \"width\": 22.459512824808826, \n                \"x\": 3230.287955666891, \n                \"y\": 1709.695754046528\n            }, \n            {\n                \"class\": \"pup\", \n                \"height\": 24.283655282917607, \n                \"type\": \"rect\", \n                \"width\": 20.875422962507855, \n                \"x\": 3009.469138921581, \n                \"y\": 1819.9960590986684\n            }, \n            {\n                \"class\": \"female\", \n                \"height\": 88.61404033064673, \n                \"type\": \"rect\", \n                \"width\": 36.2124684043506, \n                \"x\": 2975.8128447575373, \n                \"y\": 1664.0694304399342\n            }, \n            {\n                \"class\": \"female\", \n                \"height\": 43.02893304516988, \n                \"type\": \"rect\", \n                \"width\": 55.809804246705426, \n                \"x\": 3027.788387643782, \n                \"y\": 1645.324152677682\n            }, \n\nLooking forward to collaborating. Hit me up, thanks!",
    "183882": "Hi Chan,\nI already did the bounding boxes annotations using a simple heuristic, considering the same box size per class. It works decent and straight-forward. Those are my delta values from the center:\ndeltas = {'pups': {'x': -12, 'y': 12},\n                       'juveniles': {'x': -40, 'y': 40},\n                       'subadult_males': {'x': -30, 'y': 30},\n                       'adult_females': {'x': -25, 'y': 25},\n                       'adult_males': {'x': -30, 'y': 30}}",
    "183992": "Have you looked into using Amazon Mechanical Turk? That could be the way to go with this.",
    "183996": "I am interested in collaborating.",
    "184379": "The problem is the images are taken from different altitudes. In some cases, adult males are only 60 pixels, and in some images, adult male is over 210 pixels end to end.",
    "184380": "I think that'd violate the contest rules.",
    "184741": "I am interested",
    "184941": "Hi guys, sorry didn't login the last few days. Can you please add me on LinkedIn (contact on my profile) and we can discuss more? I am not able to send messages here.\n\nAlso, please try installing and using sloth. It took me a bit of time setting it up. If you manage to do it, I can pass you the config file I wrote for this dataset.",
    "186172": "I am interested as well",
    "186360": "I don't think that it matters much. In the end the score regards the number of elements per class not about predicting the right rectangle. An average size can be good enough.",
    "193870": "If you are using Mac OS X, you can use RectLabel. \n\nAn image annotation tool to label images for bounding box object detection and segmentation.\n\nhttps://rectlabel.com\n\nKey features:\n\n- Drawing bounding box, polygon, and cubic bezier\n\n- Export index color mask image and separated mask images\n\n- 1-click buttons make your labeling work faster\n\n- Customize the label dialog to combine with attributes\n\n- Settings for objects, attributes, hotkeys, and labeling fast\n\n- Layer order for overlapped boxes\n\n- Quick zoom to existing boxes\n\n- Support the PASCAL VOC format"
  },
  "source": "meta"
}