{
  "id": 303565,
  "title": "Any other way to avoid missing cots rather than high resolution?",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/303565",
  "author_name": "",
  "post_date": "2022-01-28T08:15:38.294280600Z",
  "votes": 8,
  "comment_count": 4,
  "views": 0,
  "content": "<p>I have try to implement such a pipline but seems get not so good result </p>\n<ol>\n<li>slices 720 1280 images to 7 * 7 smaller 180 320 images with overlap 0.5 (with sliced bboxes)</li>\n<li>clearn up  small bboxes under 0.2 orignal size and below 64px</li>\n<li>build a efficientnet classifier with a small mask aux loss, to classify each small images has cots<br>\nthis part is also with carefullt data auguments with remove small bbox<br>\nthis part is result in about 88% recall (with BCE positive weight 0.9 negtive weight 0.1)<br>\nthis recall is really not enough</li>\n<li>use a yolov5m6 to detect bbox with flip mixup and anchor recalculate (width 640px)</li>\n</ol>\n<p>but still get very poor result 0.533</p>\n<p>I'm also try to add p2 feature to detector with still seems not significante changes</p>\n<p>I visuallize result find according to bbox loss , give a proper bbox is easy for model , but there are aways so mamy missing as obj_loss is high </p>\n<p>Does any body give me some suggestion or resources for me to keep investigation? </p>\n<hr>\n<p>a visualize of pred result <br>\ngreen box is visualize of classifer produced slices<br>\nred box is preds</p>\n<p><img src=\"https://www.kaggleusercontent.com/kf/86286134/eyJhbGciOiJkaXIiLCJlbmMiOiJBMTI4Q0JDLUhTMjU2In0..iPbnYUEL87qGAVurHt9WZA.wRTKlgj-apVvXm3oICg0HCRs9HPTw64c50rniIvR6Kj-3CialfYq8bUy2PB4nfSYKOe1emLqkSUg6cP0tp7Q2ZxPrHUpslWBkMIN1Ir-WcxMj4g-cuDwFw_m5NR448FeWG597JKfzKC3pwT0RaeOaC1pQ-tybuxqk7jq7MJ06SCRDRg4SmmJqmC7iaIoUWMJ4AXGhQm6oTRaZ1T1PqTJ7HlcxU9_6pDn-HSOjP2pUydQHO4Jg0bAYxXJzPj03pBT66_KnKY7HWZVT1nlWdvFrIXWLwVvagAHfq7Y9u6oNXygnwsaZ8MSJ02OiIuhkn4ud--hivdvSv249qkErEQ5gfFSZoiB6mU24m8MFfjCD-G66GCYSFvJLopetSx6MG6A7jGq2Wpt1Qbj_E59a4suCVS4qMnocWKeMgNKd8JOr-pnUNQde7LTJnYfEt-nqEdPIOpaPNP5lcgm6l1odZu3viC_N1QjHdcvDjxXFbjnvDcw3oPhJXWcEnpXPYP_5Sy8SVDFMGnCFnNj2BrTMkmK-QHrsbb2EV7nQSORfIv5i72NwzeLDnWk6fgBtZ14P0HC7s37R7wPfsJVYLWHWuII26Xzr_yBGMYM41cerxV6rBd8_MVym3yJrZu0i7hzv7fwdrXDg_uDg9v-2F2LSMXO9w.UZeRtp9t161MIbAjFOQMAg/__results___files/__results___20_1.png\" alt=\"\"></p>\n<hr>\n<p>inference code is here </p>\n<p><a href=\"https://www.kaggle.com/drzhuzhe/slices-submit\" target=\"_blank\">https://www.kaggle.com/drzhuzhe/slices-submit</a></p>",
  "messages": [
    {
      "id": "1666983",
      "postDate": "01/28/2022 08:15:38",
      "content": "<p>I have try to implement such a pipline but seems get not so good result </p>\n<ol>\n<li>slices 720 1280 images to 7 * 7 smaller 180 320 images with overlap 0.5 (with sliced bboxes)</li>\n<li>clearn up  small bboxes under 0.2 orignal size and below 64px</li>\n<li>build a efficientnet classifier with a small mask aux loss, to classify each small images has cots<br>\nthis part is also with carefullt data auguments with remove small bbox<br>\nthis part is result in about 88% recall (with BCE positive weight 0.9 negtive weight 0.1)<br>\nthis recall is really not enough</li>\n<li>use a yolov5m6 to detect bbox with flip mixup and anchor recalculate (width 640px)</li>\n</ol>\n<p>but still get very poor result 0.533</p>\n<p>I'm also try to add p2 feature to detector with still seems not significante changes</p>\n<p>I visuallize result find according to bbox loss , give a proper bbox is easy for model , but there are aways so mamy missing as obj_loss is high </p>\n<p>Does any body give me some suggestion or resources for me to keep investigation? </p>\n<hr>\n<p>a visualize of pred result <br>\ngreen box is visualize of classifer produced slices<br>\nred box is preds</p>\n<p><img src=\"https://www.kaggleusercontent.com/kf/86286134/eyJhbGciOiJkaXIiLCJlbmMiOiJBMTI4Q0JDLUhTMjU2In0..iPbnYUEL87qGAVurHt9WZA.wRTKlgj-apVvXm3oICg0HCRs9HPTw64c50rniIvR6Kj-3CialfYq8bUy2PB4nfSYKOe1emLqkSUg6cP0tp7Q2ZxPrHUpslWBkMIN1Ir-WcxMj4g-cuDwFw_m5NR448FeWG597JKfzKC3pwT0RaeOaC1pQ-tybuxqk7jq7MJ06SCRDRg4SmmJqmC7iaIoUWMJ4AXGhQm6oTRaZ1T1PqTJ7HlcxU9_6pDn-HSOjP2pUydQHO4Jg0bAYxXJzPj03pBT66_KnKY7HWZVT1nlWdvFrIXWLwVvagAHfq7Y9u6oNXygnwsaZ8MSJ02OiIuhkn4ud--hivdvSv249qkErEQ5gfFSZoiB6mU24m8MFfjCD-G66GCYSFvJLopetSx6MG6A7jGq2Wpt1Qbj_E59a4suCVS4qMnocWKeMgNKd8JOr-pnUNQde7LTJnYfEt-nqEdPIOpaPNP5lcgm6l1odZu3viC_N1QjHdcvDjxXFbjnvDcw3oPhJXWcEnpXPYP_5Sy8SVDFMGnCFnNj2BrTMkmK-QHrsbb2EV7nQSORfIv5i72NwzeLDnWk6fgBtZ14P0HC7s37R7wPfsJVYLWHWuII26Xzr_yBGMYM41cerxV6rBd8_MVym3yJrZu0i7hzv7fwdrXDg_uDg9v-2F2LSMXO9w.UZeRtp9t161MIbAjFOQMAg/__results___files/__results___20_1.png\" alt=\"\"></p>\n<hr>\n<p>inference code is here </p>\n<p><a href=\"https://www.kaggle.com/drzhuzhe/slices-submit\" target=\"_blank\">https://www.kaggle.com/drzhuzhe/slices-submit</a></p>",
      "rawMarkdown": "I have try to implement such a pipline but seems get not so good result \n\n1. slices 720 1280 images to 7 * 7 smaller 180 320 images with overlap 0.5 (with sliced bboxes)\n2. clearn up  small bboxes under 0.2 orignal size and below 64px\n3. build a efficientnet classifier with a small mask aux loss, to classify each small images has cots\n this part is also with carefullt data auguments with remove small bbox\n this part is result in about 88% recall (with BCE positive weight 0.9 negtive weight 0.1)\n this recall is really not enough\n4. use a yolov5m6 to detect bbox with flip mixup and anchor recalculate (width 640px)\n\nbut still get very poor result 0.533\n\n\nI'm also try to add p2 feature to detector with still seems not significante changes\n\nI visuallize result find according to bbox loss , give a proper bbox is easy for model , but there are aways so mamy missing as obj_loss is high \n\nDoes any body give me some suggestion or resources for me to keep investigation? \n\n\n\n----------------------------------------\n\n\na visualize of pred result \ngreen box is visualize of classifer produced slices\nred box is preds\n\n\n![](https://www.kaggleusercontent.com/kf/86286134/eyJhbGciOiJkaXIiLCJlbmMiOiJBMTI4Q0JDLUhTMjU2In0..iPbnYUEL87qGAVurHt9WZA.wRTKlgj-apVvXm3oICg0HCRs9HPTw64c50rniIvR6Kj-3CialfYq8bUy2PB4nfSYKOe1emLqkSUg6cP0tp7Q2ZxPrHUpslWBkMIN1Ir-WcxMj4g-cuDwFw_m5NR448FeWG597JKfzKC3pwT0RaeOaC1pQ-tybuxqk7jq7MJ06SCRDRg4SmmJqmC7iaIoUWMJ4AXGhQm6oTRaZ1T1PqTJ7HlcxU9_6pDn-HSOjP2pUydQHO4Jg0bAYxXJzPj03pBT66_KnKY7HWZVT1nlWdvFrIXWLwVvagAHfq7Y9u6oNXygnwsaZ8MSJ02OiIuhkn4ud--hivdvSv249qkErEQ5gfFSZoiB6mU24m8MFfjCD-G66GCYSFvJLopetSx6MG6A7jGq2Wpt1Qbj_E59a4suCVS4qMnocWKeMgNKd8JOr-pnUNQde7LTJnYfEt-nqEdPIOpaPNP5lcgm6l1odZu3viC_N1QjHdcvDjxXFbjnvDcw3oPhJXWcEnpXPYP_5Sy8SVDFMGnCFnNj2BrTMkmK-QHrsbb2EV7nQSORfIv5i72NwzeLDnWk6fgBtZ14P0HC7s37R7wPfsJVYLWHWuII26Xzr_yBGMYM41cerxV6rBd8_MVym3yJrZu0i7hzv7fwdrXDg_uDg9v-2F2LSMXO9w.UZeRtp9t161MIbAjFOQMAg/__results___files/__results___20_1.png)\n\n\n\n--------------------------------------------\n\ninference code is here \n\nhttps://www.kaggle.com/drzhuzhe/slices-submit",
      "votes": null
    },
    {
      "id": "1667807",
      "postDate": "01/29/2022 00:38:03",
      "content": "<p>did you try to slice with different sizes and compare their result?  different thresholds for clean data?</p>\n<p>typically you can use your model to predict the unlabelled images and automaticly/manually check/verify to produce new dataset for training.</p>",
      "rawMarkdown": "did you try to slice with different sizes and compare their result?  different thresholds for clean data?\n\ntypically you can use your model to predict the unlabelled images and automaticly/manually check/verify to produce new dataset for training.",
      "votes": null
    },
    {
      "id": "1667907",
      "postDate": "01/29/2022 04:25:20",
      "content": "<p><code>slices 720 1280 images to 7 * 7 smaller 180 320 images with overlap 0.5 (with sliced bboxes)</code><br>\nif you are training on these slices, then how are you predicting on the test data? like I think if you train on this slice and predict on the full image, your score will drop.</p>",
      "rawMarkdown": "`slices 720 1280 images to 7 * 7 smaller 180 320 images with overlap 0.5 (with sliced bboxes)`\nif you are training on these slices, then how are you predicting on the test data? like I think if you train on this slice and predict on the full image, your score will drop.",
      "votes": null
    },
    {
      "id": "1667912",
      "postDate": "01/29/2022 04:28:22",
      "content": "<p>I manually check it like this <br>\ngreen box is visualize of classifer produced slices<br>\nred box is preds<br>\n<img src=\"https://www.kaggleusercontent.com/kf/86286134/eyJhbGciOiJkaXIiLCJlbmMiOiJBMTI4Q0JDLUhTMjU2In0..iPbnYUEL87qGAVurHt9WZA.wRTKlgj-apVvXm3oICg0HCRs9HPTw64c50rniIvR6Kj-3CialfYq8bUy2PB4nfSYKOe1emLqkSUg6cP0tp7Q2ZxPrHUpslWBkMIN1Ir-WcxMj4g-cuDwFw_m5NR448FeWG597JKfzKC3pwT0RaeOaC1pQ-tybuxqk7jq7MJ06SCRDRg4SmmJqmC7iaIoUWMJ4AXGhQm6oTRaZ1T1PqTJ7HlcxU9_6pDn-HSOjP2pUydQHO4Jg0bAYxXJzPj03pBT66_KnKY7HWZVT1nlWdvFrIXWLwVvagAHfq7Y9u6oNXygnwsaZ8MSJ02OiIuhkn4ud--hivdvSv249qkErEQ5gfFSZoiB6mU24m8MFfjCD-G66GCYSFvJLopetSx6MG6A7jGq2Wpt1Qbj_E59a4suCVS4qMnocWKeMgNKd8JOr-pnUNQde7LTJnYfEt-nqEdPIOpaPNP5lcgm6l1odZu3viC_N1QjHdcvDjxXFbjnvDcw3oPhJXWcEnpXPYP_5Sy8SVDFMGnCFnNj2BrTMkmK-QHrsbb2EV7nQSORfIv5i72NwzeLDnWk6fgBtZ14P0HC7s37R7wPfsJVYLWHWuII26Xzr_yBGMYM41cerxV6rBd8_MVym3yJrZu0i7hzv7fwdrXDg_uDg9v-2F2LSMXO9w.UZeRtp9t161MIbAjFOQMAg/__results___files/__results___20_1.png\" alt=\"\"></p>\n<p>I think add all fold , scale up all fold , add wbf will boost about 0.15 to about 0.68 <br>\nbut it take too much time , now it is about 60 minites for single fold</p>",
      "rawMarkdown": "I manually check it like this \ngreen box is visualize of classifer produced slices\nred box is preds\n![](https://www.kaggleusercontent.com/kf/86286134/eyJhbGciOiJkaXIiLCJlbmMiOiJBMTI4Q0JDLUhTMjU2In0..iPbnYUEL87qGAVurHt9WZA.wRTKlgj-apVvXm3oICg0HCRs9HPTw64c50rniIvR6Kj-3CialfYq8bUy2PB4nfSYKOe1emLqkSUg6cP0tp7Q2ZxPrHUpslWBkMIN1Ir-WcxMj4g-cuDwFw_m5NR448FeWG597JKfzKC3pwT0RaeOaC1pQ-tybuxqk7jq7MJ06SCRDRg4SmmJqmC7iaIoUWMJ4AXGhQm6oTRaZ1T1PqTJ7HlcxU9_6pDn-HSOjP2pUydQHO4Jg0bAYxXJzPj03pBT66_KnKY7HWZVT1nlWdvFrIXWLwVvagAHfq7Y9u6oNXygnwsaZ8MSJ02OiIuhkn4ud--hivdvSv249qkErEQ5gfFSZoiB6mU24m8MFfjCD-G66GCYSFvJLopetSx6MG6A7jGq2Wpt1Qbj_E59a4suCVS4qMnocWKeMgNKd8JOr-pnUNQde7LTJnYfEt-nqEdPIOpaPNP5lcgm6l1odZu3viC_N1QjHdcvDjxXFbjnvDcw3oPhJXWcEnpXPYP_5Sy8SVDFMGnCFnNj2BrTMkmK-QHrsbb2EV7nQSORfIv5i72NwzeLDnWk6fgBtZ14P0HC7s37R7wPfsJVYLWHWuII26Xzr_yBGMYM41cerxV6rBd8_MVym3yJrZu0i7hzv7fwdrXDg_uDg9v-2F2LSMXO9w.UZeRtp9t161MIbAjFOQMAg/__results___files/__results___20_1.png)\n\nI think add all fold , scale up all fold , add wbf will boost about 0.15 to about 0.68 \nbut it take too much time , now it is about 60 minites for single fold",
      "votes": null
    },
    {
      "id": "1667913",
      "postDate": "01/29/2022 04:30:46",
      "content": "<p>I use a two stage algorthem </p>\n<ol>\n<li>first, I check if there are cots on slice, only classify a bool result that If there is Cots on every Slices with (180 320)</li>\n<li>second, when detect only pred on first stage filtered slices <br>\nthen gether result together and use NMS to reduce duplicate</li>\n</ol>",
      "rawMarkdown": "I use a two stage algorthem \n\n1. first, I check if there are cots on slice, only classify a bool result that If there is Cots on every Slices with (180 320)\n2. second, when detect only pred on first stage filtered slices \n then gether result together and use NMS to reduce duplicate",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1667807,
      "author_name": "dragonzhang",
      "author_url": "",
      "post_date": "01/29/2022 00:38:03",
      "content": "<p>did you try to slice with different sizes and compare their result?  different thresholds for clean data?</p>\n<p>typically you can use your model to predict the unlabelled images and automaticly/manually check/verify to produce new dataset for training.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1667912,
          "author_name": "drzhuzhe",
          "author_url": "",
          "post_date": "01/29/2022 04:28:22",
          "content": "<p>I manually check it like this <br>\ngreen box is visualize of classifer produced slices<br>\nred box is preds<br>\n<img src=\"https://www.kaggleusercontent.com/kf/86286134/eyJhbGciOiJkaXIiLCJlbmMiOiJBMTI4Q0JDLUhTMjU2In0..iPbnYUEL87qGAVurHt9WZA.wRTKlgj-apVvXm3oICg0HCRs9HPTw64c50rniIvR6Kj-3CialfYq8bUy2PB4nfSYKOe1emLqkSUg6cP0tp7Q2ZxPrHUpslWBkMIN1Ir-WcxMj4g-cuDwFw_m5NR448FeWG597JKfzKC3pwT0RaeOaC1pQ-tybuxqk7jq7MJ06SCRDRg4SmmJqmC7iaIoUWMJ4AXGhQm6oTRaZ1T1PqTJ7HlcxU9_6pDn-HSOjP2pUydQHO4Jg0bAYxXJzPj03pBT66_KnKY7HWZVT1nlWdvFrIXWLwVvagAHfq7Y9u6oNXygnwsaZ8MSJ02OiIuhkn4ud--hivdvSv249qkErEQ5gfFSZoiB6mU24m8MFfjCD-G66GCYSFvJLopetSx6MG6A7jGq2Wpt1Qbj_E59a4suCVS4qMnocWKeMgNKd8JOr-pnUNQde7LTJnYfEt-nqEdPIOpaPNP5lcgm6l1odZu3viC_N1QjHdcvDjxXFbjnvDcw3oPhJXWcEnpXPYP_5Sy8SVDFMGnCFnNj2BrTMkmK-QHrsbb2EV7nQSORfIv5i72NwzeLDnWk6fgBtZ14P0HC7s37R7wPfsJVYLWHWuII26Xzr_yBGMYM41cerxV6rBd8_MVym3yJrZu0i7hzv7fwdrXDg_uDg9v-2F2LSMXO9w.UZeRtp9t161MIbAjFOQMAg/__results___files/__results___20_1.png\" alt=\"\"></p>\n<p>I think add all fold , scale up all fold , add wbf will boost about 0.15 to about 0.68 <br>\nbut it take too much time , now it is about 60 minites for single fold</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1667907,
      "author_name": "mrinath",
      "author_url": "",
      "post_date": "01/29/2022 04:25:20",
      "content": "<p><code>slices 720 1280 images to 7 * 7 smaller 180 320 images with overlap 0.5 (with sliced bboxes)</code><br>\nif you are training on these slices, then how are you predicting on the test data? like I think if you train on this slice and predict on the full image, your score will drop.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1667913,
          "author_name": "drzhuzhe",
          "author_url": "",
          "post_date": "01/29/2022 04:30:46",
          "content": "<p>I use a two stage algorthem </p>\n<ol>\n<li>first, I check if there are cots on slice, only classify a bool result that If there is Cots on every Slices with (180 320)</li>\n<li>second, when detect only pred on first stage filtered slices <br>\nthen gether result together and use NMS to reduce duplicate</li>\n</ol>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1666983": "I have try to implement such a pipline but seems get not so good result \n\n1. slices 720 1280 images to 7 * 7 smaller 180 320 images with overlap 0.5 (with sliced bboxes)\n2. clearn up  small bboxes under 0.2 orignal size and below 64px\n3. build a efficientnet classifier with a small mask aux loss, to classify each small images has cots\n this part is also with carefullt data auguments with remove small bbox\n this part is result in about 88% recall (with BCE positive weight 0.9 negtive weight 0.1)\n this recall is really not enough\n4. use a yolov5m6 to detect bbox with flip mixup and anchor recalculate (width 640px)\n\nbut still get very poor result 0.533\n\n\nI'm also try to add p2 feature to detector with still seems not significante changes\n\nI visuallize result find according to bbox loss , give a proper bbox is easy for model , but there are aways so mamy missing as obj_loss is high \n\nDoes any body give me some suggestion or resources for me to keep investigation? \n\n\n\n----------------------------------------\n\n\na visualize of pred result \ngreen box is visualize of classifer produced slices\nred box is preds\n\n\n![](https://www.kaggleusercontent.com/kf/86286134/eyJhbGciOiJkaXIiLCJlbmMiOiJBMTI4Q0JDLUhTMjU2In0..iPbnYUEL87qGAVurHt9WZA.wRTKlgj-apVvXm3oICg0HCRs9HPTw64c50rniIvR6Kj-3CialfYq8bUy2PB4nfSYKOe1emLqkSUg6cP0tp7Q2ZxPrHUpslWBkMIN1Ir-WcxMj4g-cuDwFw_m5NR448FeWG597JKfzKC3pwT0RaeOaC1pQ-tybuxqk7jq7MJ06SCRDRg4SmmJqmC7iaIoUWMJ4AXGhQm6oTRaZ1T1PqTJ7HlcxU9_6pDn-HSOjP2pUydQHO4Jg0bAYxXJzPj03pBT66_KnKY7HWZVT1nlWdvFrIXWLwVvagAHfq7Y9u6oNXygnwsaZ8MSJ02OiIuhkn4ud--hivdvSv249qkErEQ5gfFSZoiB6mU24m8MFfjCD-G66GCYSFvJLopetSx6MG6A7jGq2Wpt1Qbj_E59a4suCVS4qMnocWKeMgNKd8JOr-pnUNQde7LTJnYfEt-nqEdPIOpaPNP5lcgm6l1odZu3viC_N1QjHdcvDjxXFbjnvDcw3oPhJXWcEnpXPYP_5Sy8SVDFMGnCFnNj2BrTMkmK-QHrsbb2EV7nQSORfIv5i72NwzeLDnWk6fgBtZ14P0HC7s37R7wPfsJVYLWHWuII26Xzr_yBGMYM41cerxV6rBd8_MVym3yJrZu0i7hzv7fwdrXDg_uDg9v-2F2LSMXO9w.UZeRtp9t161MIbAjFOQMAg/__results___files/__results___20_1.png)\n\n\n\n--------------------------------------------\n\ninference code is here \n\nhttps://www.kaggle.com/drzhuzhe/slices-submit",
    "1667807": "did you try to slice with different sizes and compare their result?  different thresholds for clean data?\n\ntypically you can use your model to predict the unlabelled images and automaticly/manually check/verify to produce new dataset for training.",
    "1667907": "`slices 720 1280 images to 7 * 7 smaller 180 320 images with overlap 0.5 (with sliced bboxes)`\nif you are training on these slices, then how are you predicting on the test data? like I think if you train on this slice and predict on the full image, your score will drop.",
    "1667912": "I manually check it like this \ngreen box is visualize of classifer produced slices\nred box is preds\n![](https://www.kaggleusercontent.com/kf/86286134/eyJhbGciOiJkaXIiLCJlbmMiOiJBMTI4Q0JDLUhTMjU2In0..iPbnYUEL87qGAVurHt9WZA.wRTKlgj-apVvXm3oICg0HCRs9HPTw64c50rniIvR6Kj-3CialfYq8bUy2PB4nfSYKOe1emLqkSUg6cP0tp7Q2ZxPrHUpslWBkMIN1Ir-WcxMj4g-cuDwFw_m5NR448FeWG597JKfzKC3pwT0RaeOaC1pQ-tybuxqk7jq7MJ06SCRDRg4SmmJqmC7iaIoUWMJ4AXGhQm6oTRaZ1T1PqTJ7HlcxU9_6pDn-HSOjP2pUydQHO4Jg0bAYxXJzPj03pBT66_KnKY7HWZVT1nlWdvFrIXWLwVvagAHfq7Y9u6oNXygnwsaZ8MSJ02OiIuhkn4ud--hivdvSv249qkErEQ5gfFSZoiB6mU24m8MFfjCD-G66GCYSFvJLopetSx6MG6A7jGq2Wpt1Qbj_E59a4suCVS4qMnocWKeMgNKd8JOr-pnUNQde7LTJnYfEt-nqEdPIOpaPNP5lcgm6l1odZu3viC_N1QjHdcvDjxXFbjnvDcw3oPhJXWcEnpXPYP_5Sy8SVDFMGnCFnNj2BrTMkmK-QHrsbb2EV7nQSORfIv5i72NwzeLDnWk6fgBtZ14P0HC7s37R7wPfsJVYLWHWuII26Xzr_yBGMYM41cerxV6rBd8_MVym3yJrZu0i7hzv7fwdrXDg_uDg9v-2F2LSMXO9w.UZeRtp9t161MIbAjFOQMAg/__results___files/__results___20_1.png)\n\nI think add all fold , scale up all fold , add wbf will boost about 0.15 to about 0.68 \nbut it take too much time , now it is about 60 minites for single fold",
    "1667913": "I use a two stage algorthem \n\n1. first, I check if there are cots on slice, only classify a bool result that If there is Cots on every Slices with (180 320)\n2. second, when detect only pred on first stage filtered slices \n then gether result together and use NMS to reduce duplicate"
  },
  "source": "meta"
}