{
  "id": 149103,
  "title": "Sub-categorization in Object Detection",
  "url": "/competitions/global-wheat-detection/discussion/149103",
  "author_name": "pablOberhauser",
  "post_date": "2020-05-06T20:54:49.147000",
  "votes": 1,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Does anyone have papers on object detection performance on a task like this one when you split up into subclasses vs. one class? </p>\n\n<p>What I mean specifically, is how do different object detectors perform if, for example, you have one class (wheat head) and a negative class (images without bounding boxes) vs. a detector that sees multiple classes like ethz_1, arvalis_1, rres_1, arvalis_3, usask_1,  arvalis_2, inrae_1, and a negative class. (Mostly thinking of this in terms of yolo)</p>\n\n<p>Perhaps some papers on hierarchical classes?</p>\n\n<p>I imagine this leads to some sampling issues since some classes have less training images?</p>\n\n<p>I think this discussion is similar to BengaliAI decision of three models vs one model with different heads, but I wonder if there have been experiments for object detection.</p>\n\n<p>I will experiment with this, but was hoping some people had some interesting papers on the subject!</p>",
  "messages": [
    {
      "id": 836259,
      "postDate": "2020-05-06T20:54:49.147Z",
      "content": "<p>Does anyone have papers on object detection performance on a task like this one when you split up into subclasses vs. one class? </p>\n\n<p>What I mean specifically, is how do different object detectors perform if, for example, you have one class (wheat head) and a negative class (images without bounding boxes) vs. a detector that sees multiple classes like ethz_1, arvalis_1, rres_1, arvalis_3, usask_1,  arvalis_2, inrae_1, and a negative class. (Mostly thinking of this in terms of yolo)</p>\n\n<p>Perhaps some papers on hierarchical classes?</p>\n\n<p>I imagine this leads to some sampling issues since some classes have less training images?</p>\n\n<p>I think this discussion is similar to BengaliAI decision of three models vs one model with different heads, but I wonder if there have been experiments for object detection.</p>\n\n<p>I will experiment with this, but was hoping some people had some interesting papers on the subject!</p>",
      "rawMarkdown": "Does anyone have papers on object detection performance on a task like this one when you split up into subclasses vs. one class? \n\nWhat I mean specifically, is how do different object detectors perform if, for example, you have one class (wheat head) and a negative class (images without bounding boxes) vs. a detector that sees multiple classes like ethz_1, arvalis_1, rres_1, arvalis_3, usask_1,  arvalis_2, inrae_1, and a negative class. (Mostly thinking of this in terms of yolo)\n\nPerhaps some papers on hierarchical classes?\n\nI imagine this leads to some sampling issues since some classes have less training images?\n\nI think this discussion is similar to BengaliAI decision of three models vs one model with different heads, but I wonder if there have been experiments for object detection.\n\nI will experiment with this, but was hoping some people had some interesting papers on the subject!\n\n\n",
      "votes": 1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "836259": "Does anyone have papers on object detection performance on a task like this one when you split up into subclasses vs. one class? \n\nWhat I mean specifically, is how do different object detectors perform if, for example, you have one class (wheat head) and a negative class (images without bounding boxes) vs. a detector that sees multiple classes like ethz_1, arvalis_1, rres_1, arvalis_3, usask_1,  arvalis_2, inrae_1, and a negative class. (Mostly thinking of this in terms of yolo)\n\nPerhaps some papers on hierarchical classes?\n\nI imagine this leads to some sampling issues since some classes have less training images?\n\nI think this discussion is similar to BengaliAI decision of three models vs one model with different heads, but I wonder if there have been experiments for object detection.\n\nI will experiment with this, but was hoping some people had some interesting papers on the subject!\n\n\n"
  }
}