{
  "id": 37240,
  "title": "Is it legal to make use of TEST SET for unsupervised learning?",
  "url": "/competitions/carvana-image-masking-challenge/discussion/37240",
  "author_name": "",
  "post_date": "2017-07-29T14:49:39.594949Z",
  "votes": 4,
  "comment_count": 3,
  "views": 0,
  "content": "<p>We all know that only the spinning platform and the cars change their pose in the images if we ignore the position variance of the camera. So we can design a model (eg. PCA, GMM) to recognize the unchanged objects as background without any labels. </p>\n\n<p>So my question is, could we train such a model on test images if there is no hand-label involved? </p>",
  "messages": [
    {
      "id": "208367",
      "postDate": "07/29/2017 14:49:39",
      "content": "<p>We all know that only the spinning platform and the cars change their pose in the images if we ignore the position variance of the camera. So we can design a model (eg. PCA, GMM) to recognize the unchanged objects as background without any labels. </p>\n\n<p>So my question is, could we train such a model on test images if there is no hand-label involved? </p>",
      "rawMarkdown": "We all know that only the spinning platform and the cars change their pose in the images if we ignore the position variance of the camera. So we can design a model (eg. PCA, GMM) to recognize the unchanged objects as background without any labels. \n\nSo my question is, could we train such a model on test images if there is no hand-label involved?",
      "votes": null
    },
    {
      "id": "208369",
      "postDate": "07/29/2017 14:54:56",
      "content": "<p>it is legal. As long as you do not use hand label, it is ok. Unsupervised learning had been used in kaggle before.\nBut i am wondering how well will robust pca and low rank methods work? you have some experimental results?</p>",
      "rawMarkdown": "it is legal. As long as you do not use hand label, it is ok. Unsupervised learning had been used in kaggle before.\nBut i am wondering how well will robust pca and low rank methods work? you have some experimental results?",
      "votes": null
    },
    {
      "id": "208382",
      "postDate": "07/29/2017 15:46:08",
      "content": "<p>Thanks for replying. I haven't try such methods yet. I'm still thinking about how to deal with the test images. PCA is just an example and may be a suboptimal choice. Hopefully some CNN-based methods such as convolutional auto-encoder may help.</p>",
      "rawMarkdown": "Thanks for replying. I haven't try such methods yet. I'm still thinking about how to deal with the test images. PCA is just an example and may be a suboptimal choice. Hopefully some CNN-based methods such as convolutional auto-encoder may help.",
      "votes": null
    },
    {
      "id": "219831",
      "postDate": "09/09/2017 21:00:58",
      "content": "<p>Hi\nI tried to cut out the background and randomly collect images from it for the training set. This data set was used both for training a single model and in the process of joint learning of combinations of models (in GAN style).</p>\n\n<p>Improved result did not observe.</p>",
      "rawMarkdown": "Hi\nI tried to cut out the background and randomly collect images from it for the training set. This data set was used both for training a single model and in the process of joint learning of combinations of models (in GAN style).\n\nImproved result did not observe.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 208369,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "07/29/2017 14:54:56",
      "content": "<p>it is legal. As long as you do not use hand label, it is ok. Unsupervised learning had been used in kaggle before.\nBut i am wondering how well will robust pca and low rank methods work? you have some experimental results?</p>",
      "votes": null,
      "replies": [
        {
          "id": 208382,
          "author_name": "dawnbreaker",
          "author_url": "",
          "post_date": "07/29/2017 15:46:08",
          "content": "<p>Thanks for replying. I haven't try such methods yet. I'm still thinking about how to deal with the test images. PCA is just an example and may be a suboptimal choice. Hopefully some CNN-based methods such as convolutional auto-encoder may help.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 219831,
      "author_name": "markpopov",
      "author_url": "",
      "post_date": "09/09/2017 21:00:58",
      "content": "<p>Hi\nI tried to cut out the background and randomly collect images from it for the training set. This data set was used both for training a single model and in the process of joint learning of combinations of models (in GAN style).</p>\n\n<p>Improved result did not observe.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "208367": "We all know that only the spinning platform and the cars change their pose in the images if we ignore the position variance of the camera. So we can design a model (eg. PCA, GMM) to recognize the unchanged objects as background without any labels. \n\nSo my question is, could we train such a model on test images if there is no hand-label involved?",
    "208369": "it is legal. As long as you do not use hand label, it is ok. Unsupervised learning had been used in kaggle before.\nBut i am wondering how well will robust pca and low rank methods work? you have some experimental results?",
    "208382": "Thanks for replying. I haven't try such methods yet. I'm still thinking about how to deal with the test images. PCA is just an example and may be a suboptimal choice. Hopefully some CNN-based methods such as convolutional auto-encoder may help.",
    "219831": "Hi\nI tried to cut out the background and randomly collect images from it for the training set. This data set was used both for training a single model and in the process of joint learning of combinations of models (in GAN style).\n\nImproved result did not observe."
  },
  "source": "meta"
}