{
  "id": 64747,
  "title": "17th place solution",
  "url": "/competitions/google-ai-open-images-object-detection-track/discussion/64747",
  "author_name": "ohnabe",
  "post_date": "2018-09-01T12:06:23.750000",
  "votes": 12,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Congrats winners and all participants!</p>\n\n<p>My solution is very simple.</p>\n\n<p><strong>First-Stage</strong> Train ResNet50-FPN Faster RCNN model with all training data (480000 iters, 8 batches). \nI trained this model to predict all 500 classes without descrimination of label's hierarchy.\nResNet50-FPN Faster RCNN code is (<a href=\"https://github.com/Hakuyume/chainer-fpn\">https://github.com/Hakuyume/chainer-fpn</a>).</p>\n\n<p><strong>Second-Stage</strong> Train this model with sampling training data. (230000 iters, 8 batches)\nI down-sampled only the classes contained in many images(&gt;= 5,000 images)\nFirst-Stage, it took a long long time with my machine, so I tried training with this down-sampled data to train rare classes well.</p>\n\n<p><strong>Third-Stage</strong> Predict test images (normal and horizontal flip).\nI added expand labels to predict result based on label's hierarchy.\ne.g.</p>\n\n<p>label, score, xmin, xmax, ymin, ymax</p>\n\n<p>FootBall Helmet, 0.5, 0.1, 0.3, 0,2, 0.4</p>\n\n<p>-&gt;</p>\n\n<p>FootBall Helmet, 0.5, 0.1, 0.3, 0,2, 0.4</p>\n\n<p>Helmet, 0.5, 0.1, 0.3, 0.2, 0.4</p>\n\n<p><a href=\"https://storage.googleapis.com/openimages/web/object_detection_metric.html\">https://storage.googleapis.com/openimages/web/object_detection_metric.html</a></p>\n\n<p><strong>Fourth-Stage</strong> combine expanded predicts (normal and horizontal filp) with non-maximum supressions. </p>",
  "messages": [
    {
      "id": 379993,
      "postDate": "2018-09-01T12:06:23.750Z",
      "content": "<p>Congrats winners and all participants!</p>\n\n<p>My solution is very simple.</p>\n\n<p><strong>First-Stage</strong> Train ResNet50-FPN Faster RCNN model with all training data (480000 iters, 8 batches). \nI trained this model to predict all 500 classes without descrimination of label's hierarchy.\nResNet50-FPN Faster RCNN code is (<a href=\"https://github.com/Hakuyume/chainer-fpn\">https://github.com/Hakuyume/chainer-fpn</a>).</p>\n\n<p><strong>Second-Stage</strong> Train this model with sampling training data. (230000 iters, 8 batches)\nI down-sampled only the classes contained in many images(&gt;= 5,000 images)\nFirst-Stage, it took a long long time with my machine, so I tried training with this down-sampled data to train rare classes well.</p>\n\n<p><strong>Third-Stage</strong> Predict test images (normal and horizontal flip).\nI added expand labels to predict result based on label's hierarchy.\ne.g.</p>\n\n<p>label, score, xmin, xmax, ymin, ymax</p>\n\n<p>FootBall Helmet, 0.5, 0.1, 0.3, 0,2, 0.4</p>\n\n<p>-&gt;</p>\n\n<p>FootBall Helmet, 0.5, 0.1, 0.3, 0,2, 0.4</p>\n\n<p>Helmet, 0.5, 0.1, 0.3, 0.2, 0.4</p>\n\n<p><a href=\"https://storage.googleapis.com/openimages/web/object_detection_metric.html\">https://storage.googleapis.com/openimages/web/object_detection_metric.html</a></p>\n\n<p><strong>Fourth-Stage</strong> combine expanded predicts (normal and horizontal filp) with non-maximum supressions. </p>",
      "rawMarkdown": "Congrats winners and all participants!\n\nMy solution is very simple.\n\n**First-Stage** Train ResNet50-FPN Faster RCNN model with all training data (480000 iters, 8 batches). \nI trained this model to predict all 500 classes without descrimination of label's hierarchy.\nResNet50-FPN Faster RCNN code is (https://github.com/Hakuyume/chainer-fpn).\n\n**Second-Stage** Train this model with sampling training data. (230000 iters, 8 batches)\nI down-sampled only the classes contained in many images(&gt;= 5,000 images)\nFirst-Stage, it took a long long time with my machine, so I tried training with this down-sampled data to train rare classes well.\n\n**Third-Stage** Predict test images (normal and horizontal flip).\nI added expand labels to predict result based on label's hierarchy.\ne.g.\n\nlabel, score, xmin, xmax, ymin, ymax\n\nFootBall Helmet, 0.5, 0.1, 0.3, 0,2, 0.4\n\n-&gt;\n\nFootBall Helmet, 0.5, 0.1, 0.3, 0,2, 0.4\n\nHelmet, 0.5, 0.1, 0.3, 0.2, 0.4\n\nhttps://storage.googleapis.com/openimages/web/object_detection_metric.html\n\n**Fourth-Stage** combine expanded predicts (normal and horizontal filp) with non-maximum supressions. ",
      "votes": 11
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "379993": "Congrats winners and all participants!\n\nMy solution is very simple.\n\n**First-Stage** Train ResNet50-FPN Faster RCNN model with all training data (480000 iters, 8 batches). \nI trained this model to predict all 500 classes without descrimination of label's hierarchy.\nResNet50-FPN Faster RCNN code is (https://github.com/Hakuyume/chainer-fpn).\n\n**Second-Stage** Train this model with sampling training data. (230000 iters, 8 batches)\nI down-sampled only the classes contained in many images(&gt;= 5,000 images)\nFirst-Stage, it took a long long time with my machine, so I tried training with this down-sampled data to train rare classes well.\n\n**Third-Stage** Predict test images (normal and horizontal flip).\nI added expand labels to predict result based on label's hierarchy.\ne.g.\n\nlabel, score, xmin, xmax, ymin, ymax\n\nFootBall Helmet, 0.5, 0.1, 0.3, 0,2, 0.4\n\n-&gt;\n\nFootBall Helmet, 0.5, 0.1, 0.3, 0,2, 0.4\n\nHelmet, 0.5, 0.1, 0.3, 0.2, 0.4\n\nhttps://storage.googleapis.com/openimages/web/object_detection_metric.html\n\n**Fourth-Stage** combine expanded predicts (normal and horizontal filp) with non-maximum supressions. "
  }
}