{
  "id": 38534,
  "title": "My worst predictions.",
  "url": "/competitions/carvana-image-masking-challenge/discussion/38534",
  "author_name": "",
  "post_date": "2017-08-24T16:52:35.799113400Z",
  "votes": 7,
  "comment_count": 6,
  "views": 0,
  "content": "<p>I simply ran this <a href=\"https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/37523\">this code.</a> and got 0.996 LB score as he said. Then, I had a look what type of images make my LB score bad.</p>\n\n<p>The below images are my 12 worst-precision images.</p>\n\n<p>id_and_scores = [('bf9932f7aca8_04', 0.99026292335115862),\n ('293a0fa72e5b_01', 0.9906091709703948),\n ('189a2a32a615_02', 0.99064661128602927),\n ('4e5ac4b9f074_01', 0.99298144116438047),\n ('bd8d5780ed04_01', 0.99313501144164762),\n ('0789bed99cb8_05', 0.99323952245476332),\n ('eeb7eeca738e_06', 0.99339836509847068),\n ('2faf504842df_09', 0.99365315793289355),\n ('bf9932f7aca8_10', 0.99376444943523345),\n ('f4cd1286d5f4_16', 0.99377898717303748),\n ('0795e132d090_14', 0.99386692880389449),\n ('98ee0624de87_09', 0.99387485172627354)]</p>\n\n<p>I guess window and dark area are difficult to detect.</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/216177/7192/worst_predictions.png\" alt=\"image\" title=\"\"></p>",
  "messages": [
    {
      "id": "216177",
      "postDate": "08/24/2017 16:52:35",
      "content": "<p>I simply ran this <a href=\"https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/37523\">this code.</a> and got 0.996 LB score as he said. Then, I had a look what type of images make my LB score bad.</p>\n\n<p>The below images are my 12 worst-precision images.</p>\n\n<p>id_and_scores = [('bf9932f7aca8_04', 0.99026292335115862),\n ('293a0fa72e5b_01', 0.9906091709703948),\n ('189a2a32a615_02', 0.99064661128602927),\n ('4e5ac4b9f074_01', 0.99298144116438047),\n ('bd8d5780ed04_01', 0.99313501144164762),\n ('0789bed99cb8_05', 0.99323952245476332),\n ('eeb7eeca738e_06', 0.99339836509847068),\n ('2faf504842df_09', 0.99365315793289355),\n ('bf9932f7aca8_10', 0.99376444943523345),\n ('f4cd1286d5f4_16', 0.99377898717303748),\n ('0795e132d090_14', 0.99386692880389449),\n ('98ee0624de87_09', 0.99387485172627354)]</p>\n\n<p>I guess window and dark area are difficult to detect.</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/216177/7192/worst_predictions.png\" alt=\"image\" title=\"\"></p>",
      "rawMarkdown": "I simply ran this [this code.][1] and got 0.996 LB score as he said. Then, I had a look what type of images make my LB score bad.\n\nThe below images are my 12 worst-precision images.\n\nid_and_scores = [('bf9932f7aca8_04', 0.99026292335115862),\n ('293a0fa72e5b_01', 0.9906091709703948),\n ('189a2a32a615_02', 0.99064661128602927),\n ('4e5ac4b9f074_01', 0.99298144116438047),\n ('bd8d5780ed04_01', 0.99313501144164762),\n ('0789bed99cb8_05', 0.99323952245476332),\n ('eeb7eeca738e_06', 0.99339836509847068),\n ('2faf504842df_09', 0.99365315793289355),\n ('bf9932f7aca8_10', 0.99376444943523345),\n ('f4cd1286d5f4_16', 0.99377898717303748),\n ('0795e132d090_14', 0.99386692880389449),\n ('98ee0624de87_09', 0.99387485172627354)]\n\nI guess window and dark area are difficult to detect.\n\n![image][2]\n\n\n  [1]: https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/37523\n  [2]: https://kaggle2.blob.core.windows.net/forum-message-attachments/216177/7192/worst_predictions.png",
      "votes": null
    },
    {
      "id": "216231",
      "postDate": "08/24/2017 22:45:39",
      "content": "<p>Your second to last one was my worst one as well.</p>",
      "rawMarkdown": "Your second to last one was my worst one as well.",
      "votes": null
    },
    {
      "id": "216250",
      "postDate": "08/25/2017 01:36:01",
      "content": "<p>for the window error, do a post processing to detect holes and fill up. For error at the edges, if you run training long enough, they well be reduced (if your network capacity is large enough). For the shadow error, i am still working on it. check your train images, you should remove the errors. the window errors may be due to human error in train label images</p>",
      "rawMarkdown": "for the window error, do a post processing to detect holes and fill up. For error at the edges, if you run training long enough, they well be reduced (if your network capacity is large enough). For the shadow error, i am still working on it. check your train images, you should remove the errors. the window errors may be due to human error in train label images",
      "votes": null
    },
    {
      "id": "216293",
      "postDate": "08/25/2017 06:12:54",
      "content": "<p>I don't think detecting holes and filling up is feasible, nobody knows what are the errors in the test set masks. Anyway, we really are \"fighting\" for 200 pixels in a million pixel image. Don't think there's much more we could do now to improve even more!</p>",
      "rawMarkdown": "I don't think detecting holes and filling up is feasible, nobody knows what are the errors in the test set masks. Anyway, we really are \"fighting\" for 200 pixels in a million pixel image. Don't think there's much more we could do now to improve even more!",
      "votes": null
    },
    {
      "id": "216363",
      "postDate": "08/25/2017 12:34:41",
      "content": "<p>My experiment shows 0.00002 improvement in lb score in hole filling</p>",
      "rawMarkdown": "My experiment shows 0.00002 improvement in lb score in hole filling",
      "votes": null
    },
    {
      "id": "216383",
      "postDate": "08/25/2017 14:12:01",
      "content": "<p>@Heng CherKeng, how do you determine that it's not the wheel spoke hole? Like down part of the mask transformed by car angle?</p>",
      "rawMarkdown": "Heng CherKeng, how do you determine that it's not the wheel spoke hole? Like down part of the mask transformed by car angle?",
      "votes": null
    },
    {
      "id": "216387",
      "postDate": "08/25/2017 14:37:55",
      "content": "<p>determine bounding  box. fill up holes only if it is within a certain limit ( x0,y0 ), (y0,y1). the limit can be different for different views. But i think holes only appears in view==2</p>",
      "rawMarkdown": "determine bounding  box. fill up holes only if it is within a certain limit ( x0,y0 ), (y0,y1). the limit can be different for different views. But i think holes only appears in view==2",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 216231,
      "author_name": "harungunaydin",
      "author_url": "",
      "post_date": "08/24/2017 22:45:39",
      "content": "<p>Your second to last one was my worst one as well.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 216250,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "08/25/2017 01:36:01",
      "content": "<p>for the window error, do a post processing to detect holes and fill up. For error at the edges, if you run training long enough, they well be reduced (if your network capacity is large enough). For the shadow error, i am still working on it. check your train images, you should remove the errors. the window errors may be due to human error in train label images</p>",
      "votes": null,
      "replies": [
        {
          "id": 216293,
          "author_name": "dvdbos",
          "author_url": "",
          "post_date": "08/25/2017 06:12:54",
          "content": "<p>I don't think detecting holes and filling up is feasible, nobody knows what are the errors in the test set masks. Anyway, we really are \"fighting\" for 200 pixels in a million pixel image. Don't think there's much more we could do now to improve even more!</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 216363,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "08/25/2017 12:34:41",
          "content": "<p>My experiment shows 0.00002 improvement in lb score in hole filling</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 216383,
          "author_name": "heyt0ny",
          "author_url": "",
          "post_date": "08/25/2017 14:12:01",
          "content": "<p>@Heng CherKeng, how do you determine that it's not the wheel spoke hole? Like down part of the mask transformed by car angle?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 216387,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "08/25/2017 14:37:55",
          "content": "<p>determine bounding  box. fill up holes only if it is within a certain limit ( x0,y0 ), (y0,y1). the limit can be different for different views. But i think holes only appears in view==2</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "216177": "I simply ran this [this code.][1] and got 0.996 LB score as he said. Then, I had a look what type of images make my LB score bad.\n\nThe below images are my 12 worst-precision images.\n\nid_and_scores = [('bf9932f7aca8_04', 0.99026292335115862),\n ('293a0fa72e5b_01', 0.9906091709703948),\n ('189a2a32a615_02', 0.99064661128602927),\n ('4e5ac4b9f074_01', 0.99298144116438047),\n ('bd8d5780ed04_01', 0.99313501144164762),\n ('0789bed99cb8_05', 0.99323952245476332),\n ('eeb7eeca738e_06', 0.99339836509847068),\n ('2faf504842df_09', 0.99365315793289355),\n ('bf9932f7aca8_10', 0.99376444943523345),\n ('f4cd1286d5f4_16', 0.99377898717303748),\n ('0795e132d090_14', 0.99386692880389449),\n ('98ee0624de87_09', 0.99387485172627354)]\n\nI guess window and dark area are difficult to detect.\n\n![image][2]\n\n\n  [1]: https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/37523\n  [2]: https://kaggle2.blob.core.windows.net/forum-message-attachments/216177/7192/worst_predictions.png",
    "216231": "Your second to last one was my worst one as well.",
    "216250": "for the window error, do a post processing to detect holes and fill up. For error at the edges, if you run training long enough, they well be reduced (if your network capacity is large enough). For the shadow error, i am still working on it. check your train images, you should remove the errors. the window errors may be due to human error in train label images",
    "216293": "I don't think detecting holes and filling up is feasible, nobody knows what are the errors in the test set masks. Anyway, we really are \"fighting\" for 200 pixels in a million pixel image. Don't think there's much more we could do now to improve even more!",
    "216363": "My experiment shows 0.00002 improvement in lb score in hole filling",
    "216383": "Heng CherKeng, how do you determine that it's not the wheel spoke hole? Like down part of the mask transformed by car angle?",
    "216387": "determine bounding  box. fill up holes only if it is within a certain limit ( x0,y0 ), (y0,y1). the limit can be different for different views. But i think holes only appears in view==2"
  },
  "source": "meta"
}