{
  "id": 33259,
  "title": "Crowded Sea Lions",
  "url": "/competitions/noaa-fisheries-steller-sea-lion-population-count/discussion/33259",
  "author_name": "",
  "post_date": "2017-05-19T17:40:55.712467400Z",
  "votes": 3,
  "comment_count": 4,
  "views": 0,
  "content": "<p>My first try CNN detected a grand total of 8 sea lions in the following picture...\n<img src=\"http://hornsey.chessclubs.org.uk/crowd.jpg\" alt=\"Do these guys really deserve so much attention?\" title=\"\"></p>\n\n<p>I have  a trained CNN that I am using to generate activation maps on 1000x1000 picture segments. Then I am using blob detection to detect individual sea lions. It works great when everyone is spread out, but on a picture like this I just get big fat blobs with no delineation between sea lions.</p>\n\n<p>I am guessing that a few people might be stuck at the same point as me.</p>\n\n<p>Can anyone suggest any papers that might suggest how we can single out each sea lion?</p>\n\n<p>I am looking for a solution that actually detects boundaries between animals, otherwise I will have to guess the number of sea lions from the size of the blob, and that would rather take the fun out of it!</p>",
  "messages": [
    {
      "id": "183904",
      "postDate": "05/19/2017 17:40:55",
      "content": "<p>My first try CNN detected a grand total of 8 sea lions in the following picture...\n<img src=\"http://hornsey.chessclubs.org.uk/crowd.jpg\" alt=\"Do these guys really deserve so much attention?\" title=\"\"></p>\n\n<p>I have  a trained CNN that I am using to generate activation maps on 1000x1000 picture segments. Then I am using blob detection to detect individual sea lions. It works great when everyone is spread out, but on a picture like this I just get big fat blobs with no delineation between sea lions.</p>\n\n<p>I am guessing that a few people might be stuck at the same point as me.</p>\n\n<p>Can anyone suggest any papers that might suggest how we can single out each sea lion?</p>\n\n<p>I am looking for a solution that actually detects boundaries between animals, otherwise I will have to guess the number of sea lions from the size of the blob, and that would rather take the fun out of it!</p>",
      "rawMarkdown": "My first try CNN detected a grand total of 8 sea lions in the following picture...\n![Do these guys really deserve so much attention?][1]\n\n\n  [1]: http://hornsey.chessclubs.org.uk/crowd.jpg\n\nI have  a trained CNN that I am using to generate activation maps on 1000x1000 picture segments. Then I am using blob detection to detect individual sea lions. It works great when everyone is spread out, but on a picture like this I just get big fat blobs with no delineation between sea lions.\n\nI am guessing that a few people might be stuck at the same point as me.\n\nCan anyone suggest any papers that might suggest how we can single out each sea lion?\n\nI am looking for a solution that actually detects boundaries between animals, otherwise I will have to guess the number of sea lions from the size of the blob, and that would rather take the fun out of it!",
      "votes": null
    },
    {
      "id": "183913",
      "postDate": "05/19/2017 18:42:40",
      "content": "<p>Here's what I currently get for that same region. (it's Train/2.jpg) I'm training a cnn to predict blobs centered on each sea lion. The predicted blobs aren't perfectly separated, but maybe it's good enough to single out sea lions. However, this might be unnecessary effort. I was thinking of doing some sort of regression between sea lion numbers and blob density. </p>\n\n<p>My major problems are that 1) I get a lot of false positives (such as the strong blob lower left. It's a rock.)  2) My network is slow. (It takes about 1 min to predict on 1 image. The test set is 18636 images. You do the math)</p>",
      "rawMarkdown": "Here's what I currently get for that same region. (it's Train/2.jpg) I'm training a cnn to predict blobs centered on each sea lion. The predicted blobs aren't perfectly separated, but maybe it's good enough to single out sea lions. However, this might be unnecessary effort. I was thinking of doing some sort of regression between sea lion numbers and blob density. \n\nMy major problems are that 1) I get a lot of false positives (such as the strong blob lower left. It's a rock.)  2) My network is slow. (It takes about 1 min to predict on 1 image. The test set is 18636 images. You do the math)",
      "votes": null
    },
    {
      "id": "183969",
      "postDate": "05/19/2017 21:24:26",
      "content": "",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "183971",
      "postDate": "05/19/2017 21:48:43",
      "content": "<p>Thanks 3+1, your picture has given me hope. If you only have to worry about the odd rock you're doing pretty well, I would say.</p>\n\n<p>It strikes me that your problems 1 and 2 can both be mitigated by using a fast pass that looks for candidate regions and then zooming in to take a slower pass at regions of interest.</p>\n\n<p>I haven't finished writing my code to do a whole image, yet, but I am aiming for a first pass that takes 5 secs per image. That's 24 hours. And if it spits out 18000 grey scale images, they can be worked on separately for more detailed analysis then I can live with that.</p>\n\n<p>I showed the pictures to my son who with child-like innocence said 'why don't you detect the lines between the seals?' I can't help feeling he has a point. Especially if we are going to count the number of different types of seals.</p>",
      "rawMarkdown": "Thanks 3+1, your picture has given me hope. If you only have to worry about the odd rock you're doing pretty well, I would say.\n\nIt strikes me that your problems 1 and 2 can both be mitigated by using a fast pass that looks for candidate regions and then zooming in to take a slower pass at regions of interest.\n\nI haven't finished writing my code to do a whole image, yet, but I am aiming for a first pass that takes 5 secs per image. That's 24 hours. And if it spits out 18000 grey scale images, they can be worked on separately for more detailed analysis then I can live with that.\n\nI showed the pictures to my son who with child-like innocence said 'why don't you detect the lines between the seals?' I can't help feeling he has a point. Especially if we are going to count the number of different types of seals.",
      "votes": null
    },
    {
      "id": "185011",
      "postDate": "05/23/2017 19:27:25",
      "content": "<p>@JamesEverard, \nI think that what your son is referring to is described in the following paper:  <a href=\"https://arxiv.org/abs/1505.04597\">https://arxiv.org/abs/1505.04597</a> describing Unet also used for satellite pictures, which algorithm already won several Kaggle competition.\nHowever, the manual labels of the contours of the sea lions in this competition is a lot of work.</p>",
      "rawMarkdown": "JamesEverard, \nI think that what your son is referring to is described in the following paper:  https://arxiv.org/abs/1505.04597 describing Unet also used for satellite pictures, which algorithm already won several Kaggle competition.\nHowever, the manual labels of the contours of the sea lions in this competition is a lot of work.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 183913,
      "author_name": "threeplusone",
      "author_url": "",
      "post_date": "05/19/2017 18:42:40",
      "content": "<p>Here's what I currently get for that same region. (it's Train/2.jpg) I'm training a cnn to predict blobs centered on each sea lion. The predicted blobs aren't perfectly separated, but maybe it's good enough to single out sea lions. However, this might be unnecessary effort. I was thinking of doing some sort of regression between sea lion numbers and blob density. </p>\n\n<p>My major problems are that 1) I get a lot of false positives (such as the strong blob lower left. It's a rock.)  2) My network is slow. (It takes about 1 min to predict on 1 image. The test set is 18636 images. You do the math)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 183969,
      "author_name": "jinkos",
      "author_url": "",
      "post_date": "05/19/2017 21:24:26",
      "content": "",
      "votes": null,
      "replies": []
    },
    {
      "id": 183971,
      "author_name": "jinkos",
      "author_url": "",
      "post_date": "05/19/2017 21:48:43",
      "content": "<p>Thanks 3+1, your picture has given me hope. If you only have to worry about the odd rock you're doing pretty well, I would say.</p>\n\n<p>It strikes me that your problems 1 and 2 can both be mitigated by using a fast pass that looks for candidate regions and then zooming in to take a slower pass at regions of interest.</p>\n\n<p>I haven't finished writing my code to do a whole image, yet, but I am aiming for a first pass that takes 5 secs per image. That's 24 hours. And if it spits out 18000 grey scale images, they can be worked on separately for more detailed analysis then I can live with that.</p>\n\n<p>I showed the pictures to my son who with child-like innocence said 'why don't you detect the lines between the seals?' I can't help feeling he has a point. Especially if we are going to count the number of different types of seals.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 185011,
      "author_name": "chabir",
      "author_url": "",
      "post_date": "05/23/2017 19:27:25",
      "content": "<p>@JamesEverard, \nI think that what your son is referring to is described in the following paper:  <a href=\"https://arxiv.org/abs/1505.04597\">https://arxiv.org/abs/1505.04597</a> describing Unet also used for satellite pictures, which algorithm already won several Kaggle competition.\nHowever, the manual labels of the contours of the sea lions in this competition is a lot of work.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "183904": "My first try CNN detected a grand total of 8 sea lions in the following picture...\n![Do these guys really deserve so much attention?][1]\n\n\n  [1]: http://hornsey.chessclubs.org.uk/crowd.jpg\n\nI have  a trained CNN that I am using to generate activation maps on 1000x1000 picture segments. Then I am using blob detection to detect individual sea lions. It works great when everyone is spread out, but on a picture like this I just get big fat blobs with no delineation between sea lions.\n\nI am guessing that a few people might be stuck at the same point as me.\n\nCan anyone suggest any papers that might suggest how we can single out each sea lion?\n\nI am looking for a solution that actually detects boundaries between animals, otherwise I will have to guess the number of sea lions from the size of the blob, and that would rather take the fun out of it!",
    "183913": "Here's what I currently get for that same region. (it's Train/2.jpg) I'm training a cnn to predict blobs centered on each sea lion. The predicted blobs aren't perfectly separated, but maybe it's good enough to single out sea lions. However, this might be unnecessary effort. I was thinking of doing some sort of regression between sea lion numbers and blob density. \n\nMy major problems are that 1) I get a lot of false positives (such as the strong blob lower left. It's a rock.)  2) My network is slow. (It takes about 1 min to predict on 1 image. The test set is 18636 images. You do the math)",
    "183969": "",
    "183971": "Thanks 3+1, your picture has given me hope. If you only have to worry about the odd rock you're doing pretty well, I would say.\n\nIt strikes me that your problems 1 and 2 can both be mitigated by using a fast pass that looks for candidate regions and then zooming in to take a slower pass at regions of interest.\n\nI haven't finished writing my code to do a whole image, yet, but I am aiming for a first pass that takes 5 secs per image. That's 24 hours. And if it spits out 18000 grey scale images, they can be worked on separately for more detailed analysis then I can live with that.\n\nI showed the pictures to my son who with child-like innocence said 'why don't you detect the lines between the seals?' I can't help feeling he has a point. Especially if we are going to count the number of different types of seals.",
    "185011": "JamesEverard, \nI think that what your son is referring to is described in the following paper:  https://arxiv.org/abs/1505.04597 describing Unet also used for satellite pictures, which algorithm already won several Kaggle competition.\nHowever, the manual labels of the contours of the sea lions in this competition is a lot of work."
  },
  "source": "meta"
}