{
  "id": 165324,
  "title": "Yolo with small objects",
  "url": "/competitions/global-wheat-detection/discussion/165324",
  "author_name": "",
  "post_date": "2020-07-09T09:07:17.695851700Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hello! I'm taking part for the first time in a object detection competition. </p>\n\n<p>In several blogs I read <em>\"YOLO has difficulty detecting objects that are small and close to each other\"</em> but I noticed that submissions with Yolo have generally higher scores compared to submissions with Faster RCNN. </p>\n\n<p>So I was wondering why this is possible. Thanks for any hint!</p>",
  "messages": [
    {
      "id": "921426",
      "postDate": "07/09/2020 09:07:17",
      "content": "<p>Hello! I'm taking part for the first time in a object detection competition. </p>\n\n<p>In several blogs I read <em>\"YOLO has difficulty detecting objects that are small and close to each other\"</em> but I noticed that submissions with Yolo have generally higher scores compared to submissions with Faster RCNN. </p>\n\n<p>So I was wondering why this is possible. Thanks for any hint!</p>",
      "rawMarkdown": "Hello! I'm taking part for the first time in a object detection competition. \n\nIn several blogs I read *\"YOLO has difficulty detecting objects that are small and close to each other\"* but I noticed that submissions with Yolo have generally higher scores compared to submissions with Faster RCNN. \n\nSo I was wondering why this is possible. Thanks for any hint!",
      "votes": null
    },
    {
      "id": "921976",
      "postDate": "07/09/2020 17:38:59",
      "content": "<p>I think it's just the way the algorithm works. Since YOLO makes prediction using grids, it can identify objects better. and for RCNNs, since they use proposal networks it's not as good. I know this is not the best explanation but reading the papers would probably a way better idea about these algorithm. In my opinion, YOLO will generally do better than Proposal Networks, but if we have a lot of data then the performance might be comparable.</p>",
      "rawMarkdown": "I think it's just the way the algorithm works. Since YOLO makes prediction using grids, it can identify objects better. and for RCNNs, since they use proposal networks it's not as good. I know this is not the best explanation but reading the papers would probably a way better idea about these algorithm. In my opinion, YOLO will generally do better than Proposal Networks, but if we have a lot of data then the performance might be comparable.",
      "votes": null
    },
    {
      "id": "926921",
      "postDate": "07/13/2020 05:15:13",
      "content": "<p>yolo takes outputs from three layer 82, 94,106 so it predicts bbox with three resolutions small, medium and large.For predicting bbox in 94th  layer it also takes output from 106th layer and same for 82 so it preserves its high level feature by doing this that's why it is good at detecting small objects.</p>",
      "rawMarkdown": "yolo takes outputs from three layer 82, 94,106 so it predicts bbox with three resolutions small, medium and large.For predicting bbox in 94th  layer it also takes output from 106th layer and same for 82 so it preserves its high level feature by doing this that's why it is good at detecting small objects.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 921976,
      "author_name": "jonykarki",
      "author_url": "",
      "post_date": "07/09/2020 17:38:59",
      "content": "<p>I think it's just the way the algorithm works. Since YOLO makes prediction using grids, it can identify objects better. and for RCNNs, since they use proposal networks it's not as good. I know this is not the best explanation but reading the papers would probably a way better idea about these algorithm. In my opinion, YOLO will generally do better than Proposal Networks, but if we have a lot of data then the performance might be comparable.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 926921,
      "author_name": "ravi02516",
      "author_url": "",
      "post_date": "07/13/2020 05:15:13",
      "content": "<p>yolo takes outputs from three layer 82, 94,106 so it predicts bbox with three resolutions small, medium and large.For predicting bbox in 94th  layer it also takes output from 106th layer and same for 82 so it preserves its high level feature by doing this that's why it is good at detecting small objects.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "921426": "Hello! I'm taking part for the first time in a object detection competition. \n\nIn several blogs I read *\"YOLO has difficulty detecting objects that are small and close to each other\"* but I noticed that submissions with Yolo have generally higher scores compared to submissions with Faster RCNN. \n\nSo I was wondering why this is possible. Thanks for any hint!",
    "921976": "I think it's just the way the algorithm works. Since YOLO makes prediction using grids, it can identify objects better. and for RCNNs, since they use proposal networks it's not as good. I know this is not the best explanation but reading the papers would probably a way better idea about these algorithm. In my opinion, YOLO will generally do better than Proposal Networks, but if we have a lot of data then the performance might be comparable.",
    "926921": "yolo takes outputs from three layer 82, 94,106 so it predicts bbox with three resolutions small, medium and large.For predicting bbox in 94th  layer it also takes output from 106th layer and same for 82 so it preserves its high level feature by doing this that's why it is good at detecting small objects."
  },
  "source": "meta"
}