{
  "id": 74459,
  "title": "The problem is that how to identify the whales with a few samples",
  "url": "/competitions/humpback-whale-identification/discussion/74459",
  "author_name": "",
  "post_date": "2018-12-12T12:53:04.930021300Z",
  "votes": null,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Some whales has very few samples of pictures.\nThere are totally 5005 whales but 2073 whales has only one picture, 4171 whales has less than 5 pictures. It is the hard part. We can use threshold to deal with the label  'new_whale' which has 9664 pictures, but the most pictures has just few pictures, It is the problem.</p>",
  "messages": [
    {
      "id": "437756",
      "postDate": "12/12/2018 12:53:04",
      "content": "<p>Some whales has very few samples of pictures.\nThere are totally 5005 whales but 2073 whales has only one picture, 4171 whales has less than 5 pictures. It is the hard part. We can use threshold to deal with the label  'new_whale' which has 9664 pictures, but the most pictures has just few pictures, It is the problem.</p>",
      "rawMarkdown": "Some whales has very few samples of pictures.\nThere are totally 5005 whales but 2073 whales has only one picture, 4171 whales has less than 5 pictures. It is the hard part. We can use threshold to deal with the label  'new_whale' which has 9664 pictures, but the most pictures has just few pictures, It is the problem.",
      "votes": null
    },
    {
      "id": "437763",
      "postDate": "12/12/2018 13:03:34",
      "content": "<p>This is true, and it is absolutely part of the challenge; we created the reference set like this because it is a reflection of reality, one shot image recognition where we want to know if we have seen the whale before -- and if so, who? -- or if the whale is new to the dataset</p>",
      "rawMarkdown": "This is true, and it is absolutely part of the challenge; we created the reference set like this because it is a reflection of reality, one shot image recognition where we want to know if we have seen the whale before -- and if so, who? -- or if the whale is new to the dataset",
      "votes": null
    },
    {
      "id": "438663",
      "postDate": "12/14/2018 02:06:54",
      "content": "<p>I am having a hard time with overfitting in this scenario. Since so many whales have just one picture, it is hard to carve out a small subset of the training dataset for validation purposes when training the classifier... </p>",
      "rawMarkdown": "I am having a hard time with overfitting in this scenario. Since so many whales have just one picture, it is hard to carve out a small subset of the training dataset for validation purposes when training the classifier...",
      "votes": null
    },
    {
      "id": "438900",
      "postDate": "12/14/2018 11:03:16",
      "content": "<p>I used keras to build my model with resnet50(pretrained notop). But it couldn't get the better result. The top_5_acc is approximately 0.0x. </p>",
      "rawMarkdown": "I used keras to build my model with resnet50(pretrained notop). But it couldn't get the better result. The top_5_acc is approximately 0.0x.",
      "votes": null
    },
    {
      "id": "444849",
      "postDate": "12/25/2018 01:30:03",
      "content": "<p>Hi Magnus <a href=\"/tiranogreatland\">@tiranogreatland</a>! I totally agree with <a href=\"/tedcheese\">@tedcheese</a> here. I had some opportunity to work with limited datasets and just couldn't do it, there weren't a lot of values for some labels. This challenge gives us a lot of space to improve on this challenging problems! Let me know if you advance on your ideas. </p>",
      "rawMarkdown": "Hi Magnus @tiranogreatland! I totally agree with @tedcheese here. I had some opportunity to work with limited datasets and just couldn't do it, there weren't a lot of values for some labels. This challenge gives us a lot of space to improve on this challenging problems! Let me know if you advance on your ideas.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 437763,
      "author_name": "tedcheese",
      "author_url": "",
      "post_date": "12/12/2018 13:03:34",
      "content": "<p>This is true, and it is absolutely part of the challenge; we created the reference set like this because it is a reflection of reality, one shot image recognition where we want to know if we have seen the whale before -- and if so, who? -- or if the whale is new to the dataset</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 438663,
      "author_name": "zakraicik",
      "author_url": "",
      "post_date": "12/14/2018 02:06:54",
      "content": "<p>I am having a hard time with overfitting in this scenario. Since so many whales have just one picture, it is hard to carve out a small subset of the training dataset for validation purposes when training the classifier... </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 438900,
      "author_name": "zhlee2017",
      "author_url": "",
      "post_date": "12/14/2018 11:03:16",
      "content": "<p>I used keras to build my model with resnet50(pretrained notop). But it couldn't get the better result. The top_5_acc is approximately 0.0x. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 444849,
      "author_name": "jhonatansilva31415",
      "author_url": "",
      "post_date": "12/25/2018 01:30:03",
      "content": "<p>Hi Magnus <a href=\"/tiranogreatland\">@tiranogreatland</a>! I totally agree with <a href=\"/tedcheese\">@tedcheese</a> here. I had some opportunity to work with limited datasets and just couldn't do it, there weren't a lot of values for some labels. This challenge gives us a lot of space to improve on this challenging problems! Let me know if you advance on your ideas. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "437756": "Some whales has very few samples of pictures.\nThere are totally 5005 whales but 2073 whales has only one picture, 4171 whales has less than 5 pictures. It is the hard part. We can use threshold to deal with the label  'new_whale' which has 9664 pictures, but the most pictures has just few pictures, It is the problem.",
    "437763": "This is true, and it is absolutely part of the challenge; we created the reference set like this because it is a reflection of reality, one shot image recognition where we want to know if we have seen the whale before -- and if so, who? -- or if the whale is new to the dataset",
    "438663": "I am having a hard time with overfitting in this scenario. Since so many whales have just one picture, it is hard to carve out a small subset of the training dataset for validation purposes when training the classifier...",
    "438900": "I used keras to build my model with resnet50(pretrained notop). But it couldn't get the better result. The top_5_acc is approximately 0.0x.",
    "444849": "Hi Magnus @tiranogreatland! I totally agree with @tedcheese here. I had some opportunity to work with limited datasets and just couldn't do it, there weren't a lot of values for some labels. This challenge gives us a lot of space to improve on this challenging problems! Let me know if you advance on your ideas."
  },
  "source": "meta"
}