{
  "id": 20287,
  "title": "unsupervised learning to separate drivers",
  "url": "/competitions/state-farm-distracted-driver-detection/discussion/20287",
  "author_name": "",
  "post_date": "2016-04-20T15:03:42.257Z",
  "votes": null,
  "comment_count": 2,
  "views": 748,
  "content": "<p>Does any one know a good way to do an unsupervised learning to cluster images by drivers? If the drivers ID are not given, how to cluster the images by them? I feel it should be a much easier problem but don't know what is the current state of art method. Thanks</p>",
  "messages": [
    {
      "id": "115874",
      "postDate": "04/20/2016 15:03:42",
      "content": "<p>Does any one know a good way to do an unsupervised learning to cluster images by drivers? If the drivers ID are not given, how to cluster the images by them? I feel it should be a much easier problem but don't know what is the current state of art method. Thanks</p>",
      "rawMarkdown": "Does any one know a good way to do an unsupervised learning to cluster images by drivers? If the drivers ID are not given, how to cluster the images by them? I feel it should be a much easier problem but don't know what is the current state of art method. Thanks",
      "votes": null
    },
    {
      "id": "115887",
      "postDate": "04/20/2016 15:54:07",
      "content": "<p>If the drivers IDs are given, you can try reducing the dimensionality with <a href=\"https://lvdmaaten.github.io/tsne/\">t-SNE</a>, using as features the flattened images and as labels the drivers IDs, followed by any simple clustering method, such as K-Means. The nice thing about this approach is that, supposing you reduced the dimensionality to 2 or 3 dimensions, you can visually inspect the clusters formed.</p>",
      "rawMarkdown": "If the drivers IDs are given, you can try reducing the dimensionality with [t-SNE][1], using as features the flattened images and as labels the drivers IDs, followed by any simple clustering method, such as K-Means. The nice thing about this approach is that, supposing you reduced the dimensionality to 2 or 3 dimensions, you can visually inspect the clusters formed.\r\n\r\n  [1]: https://lvdmaaten.github.io/tsne/",
      "votes": null
    },
    {
      "id": "115953",
      "postDate": "04/21/2016 00:37:31",
      "content": "<p>Maybe U can first use a face detector to extract each image's face region(maybe use OpenCV?), The flatten it calculate the distances  </p>",
      "rawMarkdown": "Maybe U can first use a face detector to extract each image's face region(maybe use OpenCV?), The flatten it calculate the distances",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 115887,
      "author_name": "pennacchio",
      "author_url": "",
      "post_date": "04/20/2016 15:54:07",
      "content": "<p>If the drivers IDs are given, you can try reducing the dimensionality with <a href=\"https://lvdmaaten.github.io/tsne/\">t-SNE</a>, using as features the flattened images and as labels the drivers IDs, followed by any simple clustering method, such as K-Means. The nice thing about this approach is that, supposing you reduced the dimensionality to 2 or 3 dimensions, you can visually inspect the clusters formed.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 115953,
      "author_name": "meanku",
      "author_url": "",
      "post_date": "04/21/2016 00:37:31",
      "content": "<p>Maybe U can first use a face detector to extract each image's face region(maybe use OpenCV?), The flatten it calculate the distances  </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "115874": "Does any one know a good way to do an unsupervised learning to cluster images by drivers? If the drivers ID are not given, how to cluster the images by them? I feel it should be a much easier problem but don't know what is the current state of art method. Thanks",
    "115887": "If the drivers IDs are given, you can try reducing the dimensionality with [t-SNE][1], using as features the flattened images and as labels the drivers IDs, followed by any simple clustering method, such as K-Means. The nice thing about this approach is that, supposing you reduced the dimensionality to 2 or 3 dimensions, you can visually inspect the clusters formed.\r\n\r\n  [1]: https://lvdmaaten.github.io/tsne/",
    "115953": "Maybe U can first use a face detector to extract each image's face region(maybe use OpenCV?), The flatten it calculate the distances"
  },
  "source": "meta"
}